Journal of Experimental Psychology:Learning, Memory, and Cognition1997, Vol. 23, No. 3,638-658

Copyright 1997 by the American Psychological Association, Inc.0278-7393/97/$3.00

Event Category Learning

Alan W. Kersten and Dorrit BillmanGeorgia Institute of Technology

This research investigated the learning of event categories, in particular, categories of simpleanimated events, each involving a causal interaction between 2 characters. Four experimentsexamined whether correlations among attributes of events are easier to learn when they formpart of a rich correlational structure than when they are independent of other correlations.Event attributes (e.g., state change, path of motion) were chosen to reflect distinctions made byverbs. Participants were presented with an unsupervised learning task and were then tested onwhether the organization of correlations affected learning. Correlations forming part of asystem of correlations were found to be better learned than isolated correlations. This findingof facilitation from correlational structure is explained in terms of a model that generatesinternal feedback to adjust the salience of attributes. These experiments also provide evidenceregarding the role of object information in events, suggesting that this role is mediated byobject category representations.

Events unfolding over time have regularity and structurejust as do the enduring objects participating in those events.Adapting to a dynamic world requires not only knowledgeof objects but also knowledge of the events in which thoseobjects participate. Capturing this knowledge in eventcategories requires a highly complex representation becauseevents can often be decomposed into a number of smaller yetmeaningful spatial entities (i.e., objects) as well as temporalentities (i.e., subevents). Unlike object knowledge, thiscomplex event knowledge must often be acquired in anunsupervised context because parents seldom label eventsfor children while the events are occurring (Tomasello &Kruger, 1992). Both children and adults, however, manageto acquire scriptlike knowledge of "what happens" inparticular situations (Nelson, L986; Schank & Abelson,1977), allowing them to anticipate future events on the basisof the present situation. How people are able to learn suchevent categories in the absence of supervision represents aserious challenge to models of concept learning, which aregenerally designed around the learning of object categoriesin a supervised context.

In the present experiments we explored the unsupervisedlearning of event categories. Our interest is in unsupervisedlearning because we believe that a primary goal of category

Alan W. Kersten and Dorrit Billman, School of Psychology,Georgia Institute of Technology.

Preliminary results from the first two experiments were reportedat the 14th Annual Conference of the Cognitive Science Society.We thank Julie Earles, Chris Hertzog, Tim Salthouse, and TonySimon for comments on earlier versions of this article.

Correspondence concerning this article should be addressed toAlan W. Kersten, who is now at the Department of Psychology,Indiana University, Bloomington, Indiana 47405-1301, or to DorritBillman, School of Psychology, Georgia Institute of Technology,Atlanta, Georgia 30332-0170. Electronic mail may be sent viaInternet to Alan W. Kersten at akersten@indiana.edu, or to DorritBillman at dorrit.billman@psych.gatech.edu. Examples of theevents used as stimuli in these experiments are accessible via theWorld Wide Web at http://php.ucs.indiana.edu/-akersten.

learning is to capture predictive structure in the world. Goodcategories allow many inferences and not simply the predic-tion of a label. We believe that much natural categorylearning occurs in the absence of supervision, particularlywhen people are learning about events. Furthermore, be-cause unsupervised learning tasks are less directive andprovide fewer constraints as to what is to be learned,studying event category learning in an unsupervised contextmay be more likely to reveal learning biases that are uniqueto events.

Rather than studying complex extended events, we de-cided to focus on a much simpler event type, namely simplecausal interactions between two objects (e.g., collisions).Causal interactions have been argued to be "prototypical"events (Slobin, 1981) and thus findings here may generalizeto other event types. Causal interactions are also importantin their own right, as indicated by studies of Language andperception. For example, Slobin has noted that childrenconsistently encode causal interactions in grammatical tran-sitive sentences earlier than other event types. Michotte(1946/1963) has further demonstrated that adults perceivecausality between projected figures even when they knowthere is no true contact. Human infants as young as 6 monthsof age have also been shown to perceive causality (Leslie &Keeble, 1987). To account for these results, Leslie (1988)has proposed that humans are born with a module respon-sible for the perception of causality, with the products of thismodule serving as the foundation for later causal reasoning.Thus, people may understand complex everyday events interms of simple causal interactions.

Two Hypotheses for the Learning of Event Categories

In this research we contrasted two hypotheses as to howevent categories are learned. One hypothesis is based ontheories of object category structure and learning. Accordingto this hypothesis, the same general principles should apply

638

EVENT CATEGORIES 639

when learning event categories as when learning objectcategories. The second hypothesis is derived from theoriesas to the structure of a certain type of event category, namelymotion verb meanings. According to this hypothesis, eventcategories are structured quite differently from object catego-ries, and thus different principles apply to their learning.

The first hypothesis assumes that although events mayinvolve quite different attributes from objects, the samestructural principles may apply when forming categoriesbased on event attributes as when forming object categories.The specific claim whose applicability to event categorylearning we tested in this work is Rosen, Mervis, Gray,Johnson, and Boyes-Braem's (1976) theory that "good"categories tend to form around rich correlational structure.Correlational structure refers to the co-occurrence of sets ofproperties in an environment In an environment with richcorrelational structure, some sets of properties are foundtogether often, while others rarely or never co-occur. Thus,given one of those co-occurring properties, one can predictthat the others will also be present. For example, beaks areoften accompanied by wings because they are found to-gether on birds, while beaks and fur rarely co-occur. On thebasis of one's category of birds, then, one can predict thatwhen an object is known to have a beak, it will also havewings. Studies of natural object categories (e.g., Malt &Smith, 1984) have demonstrated that people are indeedsensitive to such correlations among properties.

Rosch et al.'s (1976) theory has implications not only forcategory structure but also for category learning mecha-nisms. That is, these learning mechanisms must be capableof detecting rich correlational structure when it is present inthe environment. More specifically, Billman and Heit (1988)have proposed that people are biased to learn correlationsforming part of a rich correlational structure and as a resultare more likely to discover a correlation when the attributesparticipating in that correlation are related to further at-tributes. In support of this theory, Billman and Knutson(1996) demonstrated that people were more likely to dis-cover a correlation between the values of two objectattributes, such as the head and tail of a novel animal, whenthose attributes were related to further attributes such asbody texture and the time of day in which the animalappeared.

There is also some evidence that the learning of eventcategories is facilitated by correlational structure, providingsupport for the hypothesis that event category learningproceeds similarly to object category learning. This evi-dence comes from work on verb learning. Although adetailed description of an event requires a complete sentencerather than just a verb, verb meanings in isolation may maponto schematic event categories. Verbs often convey informa-tion about the paths or the manners of motion of objects(Talmy, 1985). Moreover, verbs may also provide informa-tion about the identities of the objects carrying out thosemotions, such as through restrictions on the number and typeof nouns allowed by a particular verb (e.g., to push requirestwo nouns, at least one of which must be able to play the roleof agent). Thus, verb meanings may reflect simple, albeithighly schematic, event categories, and principles that apply

to the acquisition of verb meanings may be relevant to thelearning of event categories in general.

Evidence for facilitated learning of richly structured eventcategories comes from work on the acquisition of instrumentverbs, such as to saw or to hammer. Such verbs seem toinvolve rich correlational structure, specifying not only theuse of a particular instrument but also particular actions andresults. For example, the verb to saw implies not only the useof a saw but also a sawing motion and the result of theaffected object being cut. Huttenlocher, Smiley, and Chamey(1983) have provided evidence for facilitated learning ofinstrument verbs. They demonstrated better comprehensionin young children for "verbs that involve highly associatedobjects" (p. 82) than for verbs matched in complexity that donot implicate a particular object.

Behrend (1990) has also provided evidence for facilitatedlearning of instrument verbs. He found that when severaldifferent verbs could apply to an event, the first verbs usedby both children and adults to describe the event were moreoften instrument verbs than verbs that describe the action orresult of an event. This is surprising because instrumentverbs are relatively infrequent in English. Behrend's expla-nation for this finding was that instrument verbs conveymore information than do other verb types. Although thisexplanation centers on communication, the use of theseinfrequent verbs by young children may also reflect facili-tated learning of these verbs because of the rich correlationalstructure in their meanings.

The second hypothesis for the learning of event categoriesis that they are learned quite differently from object catego-ries. This hypothesis is suggested by the observation thatmost verb meanings, unlike instrument verb meanings, arestructured quite differently from object categories. In particu-lar, Huttenlocher and Lui (1979; see also Graesser, Hopkin-son, & Schmid, 1987) have proposed that verb meanings areorganized in a matrix. A matrix organization is one in whichdifferent attributes vary independently of one another andthus form separate bases for organizing a domain. Forexample, path and manner of motion are independentorganizing principles in the domain of motion events(Talmy, 1985), and thus more than one verb can apply to agiven motion event. For example, an event in whichsomeone runs into a building can be thought of as eitherrunning or entering.

This organization of verb meanings also has implicationsfor correlational structure. Because there exist multiple waysof classifying the same event, each basis for classificationtends to involve relatively few attributes, compared with thecase in which a dominant organizing principle is present. Forexample, verbs such as entering convey little informationbeyond path because path varies independently of otherattributes such as those involving manner of motion. Al-though path and manner may in fact each reflect a number ofrelated types of information rather than being unitaryattributes (e.g., the manner of motion of a creature mayinvolve the motion of its limbs relative to its body, the waythat the body as a whole moves along its path, etc.), thecorrelational structure found in such categories seems to be

640 KERSTEN AND BILLMAN

relatively sparse compared with that associated with acategory such as "bird"

These differences in structure between nouns and verbsmay have implications for the learning of object and eventcategories. For example, Gentner (1981) has argued that thericher correlational structure associated with object catego-ries in part accounts for the faster learning of nouns than ofverbs by most children. Gentner has proposed that nounmeanings, which generally refer to object categories, tend tobe associated with the highly intercorrelated attributes foundwithin events, namely the objects participating in thoseevents. Relational terms, such as verbs, are then associatedwith the remaining, relatively uncorrelated attributes. If thisaccount is correct, people may expect relatively weakcorrelational structure when learning verb meanings andpossibly when learning event categories in general. Theseexpectations could trigger different learning strategies in thecontext of an event category learning task than in an objectcategory learning task, resulting in little or no facilitation orpossibly even overshadowing of event correlations formingpart of a rich correlational structure.

Gentner's (1981) theory suggests an alternative explana-tion for the finding of facilitated learning of instrumentverbs. In particular, this facilitation may reflect the strongrelation of these verbs to particular objects. Not only doinstrument verbs such as to saw implicate the use of aparticular object, they often share a common word stem witha noun (i.e., a saw). Because nouns are generally easier forchildren to learn, this tight linkage of instrument verbs toobjects may help children learn what the verbs mean. Thus,it may not be necessary to appeal to correlational structure toaccount for the learning of instrument verbs.

A second difference between object and event categoriesalso favors the hypothesis that event categories should notshow facilitation from correlational structure. In particular,the fact that different information becomes available atdifferent points in an event may make unsupervised eventcategory learning more similar to supervised than to unsuper-vised object category learning. Even when no categorylabels are provided and the experimenter considers the taskto be unsupervised, participants may consider the task to beone of predicting the outcome of an event on the basis ofearlier predictor attributes. The eventual display of thisinformation would then act as feedback regarding theparticipant's predictions. Such temporal relations are similarto those found in supervised category learning, in whichfeedback in the form of a category label is often withhelduntil the end of a trial.

In contrast to unsupervised learning, the results of super-vised category learning experiments generally reveal notfacilitation but rather an overshadowing of correlationsforming part of a rich correlational structure. For example,Gluck and Bower (1988) found that participants were lesslikely to learn a symptom's predictiveness of a particulardisease if a second predictor was also present. Thus,participants were more likely to learn a correlation betweena predictor and an outcome when it was isolated than when itformed part of a richer correlational structure involving twopredictors and an outcome. Participants learning about

events may similarly consider the task to be one ofpredicting the outcome of an event, and thus may be lesslikely to learn further correlations when an adequate predic-tor of this outcome is found.

There are thus two alternative hypotheses as to the effectsof correlational structure on event category learning. Priorwork on unsupervised object category learning and real-world verb learning provides evidence for facilitated learn-ing of categories formed around rich correlational structure.Perhaps category learning for events as well as for objects isgeared toward learning richly structured categories. Theo-ries as to the structure of verb meanings, however, suggestthat most event categories may be structured differently thanobject categories. If so, event category learning may proceedquite differently from object category learning. Differencesbetween object and event category learning tasks in the wayinformation is revealed also favor this hypothesis. Still,because evidence from learning seems most relevant to thepresent research question, and this evidence suggests facili-tation from correlational structure for both object categoriesand verb meanings, we favored the first hypothesis thatevent categories would show facilitated learning with richcorrelational structure.

Overview of Experiments

In the present experiments we tested whether eventcategories with rich correlational structure are learned moreeasily than less structured categories. Although our predic-tions were motivated in part by prior work on verb learning(Huttenlocher et al., 1983), we designed our task moreclosely around prior work on unsupervised object categorylearning (Billman & Knutson, 1996). Thus, we tested forknowledge of event categories following an unsupervisedlearning task, in which no category labels were provided. Wedid this because we believe that the purpose of categoriza-tion is more general than that of communication, allowingone to predict future occurrences on the basis of a number ofcues, both verbal and nonverbal. Because predictions of thefuture are made possible by knowledge of past correlations,and a set of correlations among properties can be consideredto constitute a category, the learning of correlations can beused as an index of category learning. Thus, we measuredcategory learning by testing a participant's ability to distin-guish events that preserved correlations present duringlearning from events that broke those correlations.

Our experiments tested whether correlations betweenevent attributes are easier to learn when forming part of asystem of correlations than when isolated from other correla-tions. Of course, when learners are exposed to a system ofcorrelations, there are more correlations available to dis-cover than when they are exposed to isolated correlations,and thus the learner is more likely to discover at least onecorrelation. But if learners have a bias to learn richlystructured categories, they should show better learning ofeach individual correlation when it forms part of a system ofcorrelations than when it is found in isolation. We hypoth-esized that the property of richly structured categories that iskey to their superior learning is high value systematicity

EVENT CATEGORIES 641

(Barsalou & Billman, 1988; Billman & Knutson, 1996). Insystems of correlations with high value systematicity, anattribute that predicts the value of one other attribute alsopredicts the values of several other attributes. We believethat human categorization is geared toward learning catego-ries with high value systematicity because such categoriesallow many inferences and are thus very useful.

In the first experiment, we compared the learning ofcorrelations forming part of a rich correlational structurewith the learning of the same correlations when part of amatrix organization. The structured condition, similar to thestructured condition used by Billman and Knutson (1996) toinvestigate object categorization, involved a number ofintercorrelated attributes in a rich correlational structure.This condition was also consistent with suggestions ofBehrend (1990) as to the structure of instrument verbmeanings. The matrix condition, in turn, was similar to theorthogonal condition of Billman arid Knutson and consistentwith the matrix organization suggested for verbs by Hutten-locher and Lui (1979). In particular, each category in thematrix condition was based on a single correlation, withthree such correlations representing independent bases forcategorizing a given event. Thus, the categories in thestructured condition had higher value systematicity than didthose in the matrix condition because attributes in the matrixcondition varied independently from most others and al-lowed few predictions as a result.

As we discussed earlier, however, there is another charac-teristic of matrices that could account for greater difficulty inlearning a correlation in the matrix condition compared withthe structured condition in Experiment 1. In particular, thematrix condition involved multiple independent correlationsthat could potentially be used as the basis for categorization.It is possible that these independent correlations couldcompete for one's attention, with the discovery of onecorrelation discouraging the discovery of others. Thus,richly structured categories could be easier to learn notbecause of high value systematicity but rather because thereare no competing correlations. To better understand themechanisms underlying the advantage of richly structuredcategories, Experiment 2 compared the learning of a correla-tion forming part of a rich correlational structure with thelearning of the same correlation in a condition in which noother correlations were present. Thus, the less structuredcondition of Experiment 2 differed from that of Experiment1- in that there were no competing correlations.

In the structured conditions of Experiments 1 and 2, eachevent was representative of only one category. As wediscussed above, however, most events can be categorizedaccording to multiple, independent bases. In Experiment 3we tested whether people preferentially learn categories onthe basis of rich correlational structure even when alterna-tive bases for categorization are present. In Experiment 4 weinvestigated the generality of facilitation from correlationalstructure across different types of content. In Experiment 4we also tied the present work more closely to traditionalwork on category learning with an additional dependentmeasure involving the sorting of instances into categories.

Experiment 1

To test the effects of correlational structure on eventcategory learning, we used simple animated events asstimuli. Three frames from an example event are shown inFigure 1. Every event involved a causal interaction betweentwo characters. Within this framework, a number of at-tributes varied from event to event. We chose event at-tributes that are specified by verb meanings. For example,the change in state of the affected character was one attribute

Figure I. Three frames from an example event. In the first frame,the characters are shown in their starting locations, here with theagent on the left and the patient on the right. In the second frame,the agent has moved to the patient, causing the patient to explode.In the third frame, the remains of the patient have moved awayfrom the agent.

642 KERSTEN AND BILLMAN

because verbs such as to break and to cut convey differentstate changes.

Correlations between attributes allowed participants topredict the value of one attribute given the value of another.We presented participants in the structured condition withevents exhibiting correlations among four attributes: agentpath, manner of motion, state change, and environment (seeFigure 2). As with instrument verbs, these attributes in-cluded the actions of one object and the change in state ofanother object resulting from those actions. Unlike instru-ment verbs, these attributes were correlated not with theappearance of the causing object but rather with the environ-ment in which the event took place to ensure that partici-pants were indeed learning event categories rather thansimply categorizing the objects taking part in the events.Because the same values of the correlated attributes alwayswent together, all of the events involving one set ofco-occurring values could be considered to constitute anevent category. For example, participants in the structuredcondition could have learned a category of events takingplace on a background of squiggly lines in which an agentmoved smoothly in pursuit of a second character, causing itto explode when they came into contact (see Figure 3).

We presented participants in the matrix condition withevents exhibiting three independent correlations, each involv-ing only two attributes (see Figure 2). These correlationsoffered independent bases for categorizing the same events.Thus, the same event could be considered an example of acategory in which an agent moved smoothly on a squigglybackground, a category in which an agent continued to

Structured Condition

Manner of motion

State change

Rgent path

Enuironment

Patient path

pursue a second character after causing it to explode, or acategory in which a blue character and a yellow characterinteracted (see Figure 3). These categories were completelyunrelated, however, so that knowing the manner of motionof an object would not allow one to predict its path. Thisorganization is similar to the way the English languagecategorizes most events. In English, the verb in a sentence ismost often related to the manner of motion of the agent in anevent, whereas prepositions or verb particles are related tothe path of that agent (Jackendoff, 1987). These twocategories combine interchangeably, however, so that know-ing the manner of motion of an object (e.g., to run vs. towalk) does not allow one to predict its path (e.g., in vs. out).Thus, the matrix structure in this experiment was similar tothe organization of English relational terms, except that allcorrelations involved nonlinguistjc attributes rather thanverbal labels because of the unsupervised nature of the task.The use of three independent correlations in this conditionalso allowed us to equate the number of possible events inthe two conditions, with 81 possible events in each condi-tion.

We used each participant's knowledge of one correlation,the target rule, as the primary measure of that participant'slearning. We tested knowledge of the target rule by present-ing events in which the value of one target rule attributeeither matched or mismatched the value predicted by theother target rule attribute. Participants rated test events as tohow well they matched learning events. Knowledge of acorrelation was indicated by lower ratings for events inwhich attribute values mismatched than for those in whichthey matched. Three different target rules were used in thisexperiment to ensure that any effects of correlational struc-ture were not specific to a particular correlation. We used thesame three target rules in both conditions. Each target ruleinvolved at least one dynamic attribute, so that these ruleswere indeed different from those used in studies of objectcategorization. We predicted better learning of a target rulewhen it formed part of a rich correlational structure (i.e., inthe structured condition) than when it was independent of allother correlations (i.e., in the matrix condition).

Rgent appearance # 0 Patient appearance

Matrix Condition

Manner of motion

Method

State change #

flgent path #

Rgent appearance

Enuironment

Patient path

Patient appearance

Figure 2. Correlations seen by participants assigned the mannerof motion-environment target rule in Experiment 1. Dark linesbetween attributes indicate correlations, such that participantscould predict the value of one correlated attribute given the value ofthe other.

Participants

Thirty undergraduates at the Georgia Institute of Technologyreceived course credit for their participation in this experiment.

Stimuli

All events. A square agent and a circular patient interacted ineach event. The two characters started in motion when theparticipant pressed the mouse button. In each event, the agentmoved into contact with the patient, causing alterations in thepatient's appearance, called the state change, after which thepatient moved away from the agent. Each event lasted about 8 s,with a black screen appearing between events for 1 s.

The events varied in a number of ways- The starting position ofthe patient was chosen randomly from a region in the center of thescreen, whereas the agent started at a varying distance away along a

EVENT CATEGORIES 643

Category 1Structured Condition

Category 2 Category 3

Matrix ConditionCategory

/ 4"\\

^ ^ r-

/\

v ^ r

v ^ r

1

—r- *

/ ' ^ -I\ -.

/ O P S

' ̂ 0

1 -̂ ^

Category 2

<-•$ 0

:• •

Category 3O

•/.•/.tif:.-O 1

" o•

• • A *•

— 0 Do •o

m

m * • " . • '

• . • - 1 *^ 0 D

• o

• i

o • • »

e

1 '

: ° . •• 1

o * • »

o

C

0• •• •

• •

0 • • o

Figure 3. Schematic depictions of the three categories defined in terms of the manner ofmotion-environment target rule in Experiment 1. Each rectangle depicts a point in one event justafter the agent has come into contact with the patient. Bidirectional arrows represent the manner ofmotion of the agent, and unidirectional arrows represent agent path. The three rows under the MatrixCondition heading represent the three values of agent path and state change, which covaried with oneanother but varied independently of the target rule attributes. For example, the three rows under theCategory 1 heading of the matrix condition vary on agent path and state change, but all involve asmooth manner of motion and a squiggly background. Variation on agent color, patient color, andpatient path is not represented. Patient path varied randomly in both conditions. Agent color andpatient color also varied randomly in the structured condition, whereas they covaried with oneanother but varied independently of all other attributes in the matrix condition.

horizontal, vertical, or diagonal path. Events also varied alongseven attributes, each of which had three possible values. Theseattributes were the appearance of the agent, the appearance of thepatient, the environment, the path of the agent, the path of thepatient, the manner of motion of the agent, and the state change ofthe patient. Table 1 describes the values of these attributes.

Learning events. There were 120 learning events. Participantsin the structured condition saw events exhibiting correlationsamong four attributes: environment, agent path, manner of motion,and state change. One correlation from among these attributes waschosen to be each participant's target rule, either (a) agentpath-environment, (b) manner of motion-environment, or (c) agent

path-manner of motion. Participants in the matrix condition sawevents exhibiting correlations between three independent pairs ofattributes. One of these pairs constituted the participant's targetrule, and two other pairs were chosen from the remaining attributes.(Figures 2 and 3 depict the correlations present for participantsassigned the manner of motion-environment target rule.) Eachvalue of the correlated attributes was shown on 40 of the learningevents. Values of the remaining attributes varied randomly on eachevent.

Test events. There were 54 test events. Eighteen events testedfor knowledge of the target rule, whereas knowledge of two othercorrelations was tested in the remaining 36 events. In 9 tests of each

644 KERSTEN AND BILLMAN

Table 1Attributes, Values, and Means of Obscuring Attributes

Attributes Values Obscured by

Agent appearance

Patient appearance

Agent path (afterstate change)

Patient path (beforestate change)

State change

Agent manner ofmotion

Environment

1. Red2. Green3. Blue1. Purple2. Brown3. Yellow1. Follow patient2. Stay in place3. Retreat1. Toward agent2. Stay in place3. Away from agent1. Explode2. Shrink3. Flash

1. Smooth motion2. Forward surges3. Zig-zag1. Squiggly lines2. Ovals3. Dots

Darkening agent

Darkening patient

Depicting agent astied down afterstate change

Depicting patient astied down untilstate change

Cloud appearingover patientafter coming incontact withagent

Cloud appearingover agent

Displaying event onblank back-ground

rule, the values of the attributes in that rule were matched as theyhad been during learning and thus are called correct events. In 9other tests of that rule, these values were mismatched and thus arecalled incorrect events. The presentation order of test items wasdetermined randomly for each participant.

To ensure that participants in the structured condition could onlyuse knowledge of the rule being tested when rating an event, weobscured the two correlated attributes not participating in that rule.For example, when a participant was tested on the manner ofmotion-agent path target rule, the event was displayed on a blankbackground, and a cloud covered the patient after contact with theagent so that the participant could not use the environment or statechange when rating the event (see Table 1 for a description of howother attributes were obscured). If attributes had not been obscured,participants in the structured condition would have been able todetect an incorrect value of a target rule attribute by using not onlyknowledge of the target rule but also knowledge of the two othercorrelations involving that attribute. This test method was neces-sary because our goal was to investigate the learning of the sametarget rules in the structured and matrix conditions and not simplyto assess whether participants had learned any correlations at all.

We also obscured two attributes for test events seen by partici-pants in the matrix condition. The same two attributes wereobscured each time a particular rule was tested. One attribute camefrom each of the two rules that was not being tested in a given trial.For example, in the matrix condition, agent appearance and statechange would have been obscured when testing the target ruleinvolving manner of motion and environment. Six example eventsshown prior to testing demonstrated how attributes were to beobscured for each participant. Randomly varying attributes contin-ued to take random values during testing. All seven attributes wereeither represented by a particular value or obscured for each testevent.

Design

The two independent variables, manipulated between subjects,were correlational structure (matrix or structured) and the target

rule being tested (manner of motion-environment, manner ofmotion-agent path, or environment-agent path). The primarydependent variable was the difference between each participant'saverage rating for events involving correctly matched values of thetarget rule attributes and his or her average rating for eventsinvolving mismatched values.

Procedure

Sessions lasted approximately 45 min. We instructed participantsto work at their own pace and to ask questions if anything wasunclear. The remaining instructions were presented by the com-puter. The participant was instructed that there were two kinds ofcreatures on another planet, one of which always moved to theother and changed its appearance. Participants were instructed thatthey were to learn about the kinds of events that happen on thisplanet and that they would later be tested on how well they coulddifferentiate events that could take place on this planet from thosethat could not.

After the 120 learning events, the 6 example test events werepresented. Next, the participant was instructed to rate each of the 54test events as to "how well it fits in with" the learning events.Participants were instructed not to give an event a low rating justbecause some attributes were obscured. Participants rated each testevent on a 5-point scale by clicking on a button labeled BAD (arating of 1), one labeled GOOD (5), or one of three unlabeled buttonsbetween them (2, 3, and 4). A sixth button was labeled REPEAT,allowing the participant to view the event as many times as desired.After testing, the experimenter asked participants whether they hadnoticed any "general patterns or regularities during the learningevents." Participants who reported one correlation were encour-aged to report any others they had noticed.

Results

Table 2 displays the mean ratings of participants in thestructured and matrix conditions for events testing the targetrules, and Figure 4 depicts the difference between ratings ofcorrect events and incorrect events in each condition. Higherdifference scores indicate a better ability to differentiate thetwo types of test items. We adopted an alpha level of .05 forall analyses in this article. An analysis of variance (ANOVA)on difference scores revealed a significant effect of correla-

Table 2Target Rule Rating Accuracy in Experiment 1

Condition

StructuredAverageAP-MoMEnv-APEnv-MoM

MatrixAverageAP-MoMEnv-APEnv-MoM

Incorrectevents

M

2.511.672.293.58

3.042.333.093.71

SD

1.521.191.451.49

1.090.731.320.84

Correctevents

M

4.674.874.274.87

3.623.783.563.53

SD

0.460.150.580.30

0.850.881.110.67

Difference

M

2.163.201.981.29

0.581.440.47

-0.18

SD

1.691.251.931.54

1.481.352.010.43

Note. AP — agent path; MoM — manner of motion; Env =environment.

EVENT CATEGORIES 645

MoM-Env

Figure 4. Mean rating differences between events testing cor-rectly matched and mismatched values of the target rule attributesin Experiment 1. Higher difference scores indicate better discrimi-nation of correct and incorrect events. Error bars reflect standarderrors. AP = agent path; MoM = manner of motion; Env =environment.

tional structure, F(l , 24) = 8.18,/? < .01, MSE = 2.28, withmeans of 2.16 (SD — 1.69) in the structured condition and0.58 (SD = 1.48) in the matrix condition. There was also aneffect of target rule, F(2,24) = 7.96, p < .05, MSE = 2.28,with highest difference scores for the correlation betweenagent path and agent manner of motion. There was noevidence for an interaction, F(2, 24) < 1.

Although we assigned each participant one rule as thetarget rule for direct comparison with the other condition, wealso tested each participant for knowledge of two othercorrelations. These two nontarget rules differed across thetwo conditions. Still, because each participant was tested forknowledge of one target rule and two nontarget rules, acombined rating score can be created for each participant byaveraging across rating difference scores for these threecorrelations. Participants in the structured condition againshowed higher scores on this measure, t(2S) = 2.50, p < .05,averaging 2.20 (SD = 1.39), compared with \A0(SD = 1.00)for the matrix condition. Table 3 displays the mean ratings ofevents testing the nontarget rules in this experiment.

We also assessed participants' knowledge of the targetrules by scoring postexperimental interviews. A participantwas given 1 point for reporting each correct pairing of valuesof the target rule attributes. Because each attribute had threepossible values, the maximum possible score was 3, with 0reflecting no correct reports. Trends in interview scores werequite similar to those of target rule rating difference scores,with a correlation of .71 (p < .001) between the twomeasures. An ANOVA on interview scores, however, failedto reveal any significant effects, although the effect ofcorrelational structure approached significance, F(\, 24) =3.25, p < .09, MSE = 1.73. The structured conditionaveraged 1.27 (SD = 1.49), compared with the matrixcondition's average of 0.40 (SD = 1.06). Six participants in thestructured condition reported all three pairings of the target ruleattributes, compared with 2 participants in the matrix condition.

Discussion

Participants in this experiment showed better learning of acorrelation when it formed part of a rich correlational

structure than when it was independent of other correlations.This finding provides evidence for the hypothesis that eventcategory learning is geared toward categories with highvalue systematicity, extending earlier findings on objectcategory learning (Billman & Knutson, 1996). The existenceof correlations independent of the target rule in the matrixcondition, however, suggests an alternative account of thepresent results. A participant who noticed one of these othercorrelations first may have subsequently paid more attentionto the attributes in that correlation at the expense of otherattributes, making the target correlation more difficult todiscover. Thus, the results of this experiment could reflectfacilitation from correlational structure in the structuredcondition, competition among independent correlations inthe matrix condition, or some combination of the two. Wedesigned Experiment 2 to determine whether the advantageof richly structured categories is found even when noindependent correlations are present in the less structuredcondition.

Experiment 2

The design of Experiment 2 was quite similar to that ofExperiment 1. There was again a structured condition, inwhich four attributes were correlated for each participant.Instead of a matrix condition, however, there was in misexperiment a condition in which only the two target ruleattributes were correlated, and all other attributes variedrandomly (see Figure 5). This condition was called theisolated condition because the attributes in the target ruleconstituted a single, isolated correlation. Thus, the isolatedcondition was like the matrix condition, except that therewere no other independent correlations present to potentiallydraw attention away from the target rule attributes. If theresults of Experiment 1 were entirely due to competition

Table 3Nontarget Rule Rating Accuracy in Experiment 1

Condition

StructuredEnv-MoMMoM-SCEnv-SCEnv-APAP-SCAP-MoM

MatrixEnv-SCAP-SCAA-PAMoM-SCAA-PPPA-PP

Incorrectevents

M

1.801.762.092.492.242.56

1.581.712.242.473.473.87

SD

1.351.081.501.370.290.87

1.060.801.351.351.040.67

Correctevents

Af

4.714.584.384.424.044.16

4.334.203.914.023.443.60

SD

0.330.320.600.580.530.45

0.970.660.630.911.160.71

Difference

M

2.912.822.291.931.801.60

2.762.491.671.56

-0.02-0.27

SD

1.601.371.971.630.781.18

2.011.441.722.240.280.28

Note. Env = environment; MoM = manner of motion; SC =state change; AP = agent path; AA = agent appearance; PA =patient appearance; PP = patient path. Rules are ordered bydifficulty in each condition, with different rules in the twoconditions.

646 KERSTEN AND BILLMAN

EHample Structured Condition

Marnier of motion

State change Environment

Rgent path # ^o> Patient path

Rgent appearance # # Patient appearance

Isolated Condition

Manner of motion

State change

Rgent path : \

Environment

Patient path

Rgenl appearance Patient appearance

Figure 5. Correlations seen by participants assigned the mannerof motion-patient path target rule in Experiment 2. The top schemais only an example of what participants saw in the structuredcondition because the actual choice of attributes to covary withmanner of motion and patient path was random.

among independent correlations, the two conditions in thisexperiment should have performed equally well because noattributes covaried independently of the target rule. Wepredicted, however, that participants would show betterlearning of the target rule when it formed part of a richcorrelational structure (i.e., in the structured condition) thanwhen it was the only correlation present (i.e., in the isolatedcondition).

Method

Participants

Thirty-six undergraduates at the Georgia Institute of Technologyreceived course credit for their participation in this experiment.

Stimuli

Learning events. The correlations present in the learningevents of Experiment 2 differed from those of Experiment 1. Weused three new target rules to explore the benefits of correlationalstructure across a variety of event attributes. These were as follows:(a) state change-environment, (b) agent path-patient appearance,and (c) patient path-(agent) manner of motion. The target rule wasthe only correlation present for participants in the isolated condi-tion. In the structured condition, two other attributes also correlatedwith the target rule attributes. These attributes were randomlychosen for each participant from the set of remaining attributes.

Test events. As in Experiment 1,54 items tested for knowledgeof three different correlational rules. Eighteen items tested forknowledge of the target rule, and the remainder were filler items.On tests of the target rule, the two correlated attributes not

participating in the target rule were obscured for participants in thestructured condition. Two attributes were also obscured throughouttesting for participants in the isolated condition to make the testprocedure equally novel for both conditions. These attributes werechosen randomly for each participant from the set of uncorrelatedattributes. Filler items seen by participants in the structuredcondition tested for knowledge of two other correlations presentduring learning. Participants in the isolated condition had no basisfor evaluating filler items because only the target rule had beenpresent during learning.

Design

The two independent variables, manipulated between subjects,were the correlational structure (isolated or structured) and thetarget rule being tested (state change-environment, agent path-patient appearance, or patient path-agent manner of motion). Theprimary dependent variable was the difference between eachparticipant's average rating for events involving correctly matchedvalues of the target rule attributes and his or her average rating forevents involving mismatched values.

Procedure

The procedure in Experiment 2 was the same as in Experiment 1.

Results

Table 4 displays the mean ratings of participants in thestructured and isolated conditions for events testing thetarget rules in this experiment, and Table 5 displays ratingsof the nontarget rules by participants in the structuredcondition. Figure 6 depicts rating differences betweencorrect and incorrect events for the two conditions. AnANOVA on difference scores again revealed a significanteffect of correlational structure, F(l, 30) = 8.82, p < .01,MSE = 1.39, with means of 1.78 (SD = 1.66) in thestructured condition and 0.61 (SD = 1.54) in the isolatedcondition. There was also an effect of target rule, F(2,30) —15.83, p < .001, MSE = 1.39, with the highest differencescores for participants tested on the correlation between state

Table 4Target Rule Rating Accuracy in Experiment 2

Condition

StructuredAverageSC-EnvPA-APPP-MoM

IsolatedAverageSC-EnvPA-APPP-MoM

Incorrectevents

M

2.361.222.853.02

3.322.153.913.89

SD

1.140.351.300.49

1.421.750.770.89

Correctevents

M

4.144.614.203.61

3.934.263.653.87

SD

0.740.210.780.77

0.980.881.021.09

Difference

M

1.783.391.350.59

0.612.11

-0.26-0.02

SD

1.660.531.800.91

1.541.840.710.31

Note. SC = state change; Env = environment; PA = patientappearance; AP = agent path; PP = patient path; MoM = mannerof motion.

EVENT CATEGORIES 647

Table 5Nontarget Rule Rating Accuracy in the Structured Condition of Experiment 2

Rule

PA-SCPA-PPAP-SCPP-MoMAA-PAPP-SCAP-PPMoM-EnvMoM-SCPA-MoMAP-MoMAA-MoMAP-EnvAA-PPAA-SCAA-AP

Incorrect events

M

1.001.221.502.782.482.632.712.863.193.893.783.673.333.893.224.00

SD

0.000.39—1.721.571.691.021.01——

—0.94——

Correct events

M

5.004.114.174.894.303.983.843.814.114.674.444.113.674.223.112.33

SD

—0.47039—

0,940.890.690.940.59—.———

0.47——

Difference

M

4.002.892.672.111.821.351.130.950.920.780.660.440.340.33

-o.u-1.67

SD

—0.470.79—1.811.981.111.041.08————

0.47——

N

1221375431111211

Note. PA = patient appearance; SC = state change; PP = patient path; AP = agent path; MoM =manner of motion; AA = agent appearance; Env = environment. The number of participants testedon each rule varied because the nontarget rules were randomly selected from the correlations seen bya given participant. Dashes indicate that standard deviations were not available for some rulesbecause only 1 participant was tested on each of those rules.

change and environment. There was no evidence for aninteraction, F(2 ,30) < 1.

Participants in the structured condition (Af=1 .50 ,SD = 1.47) also performed better than participants in theisolated condition (M = 0.67, SD — 1.28) on interviewscores, F ( l , 30) = 6.82, p < .05, MSE = 0.92. Sevenparticipants in the structured condition reported the correctpairings of all three values of the target rule attributes,compared with 4 participants in the isolated condition. Therewas also a significant effect of target rule on interviewscores, F(2, 30) = 19.91, p < .001, MSE = 0.92. Interviewscores averaged 2.50 (SD = 1.17) on the correlation be-tween state change and environment, 0.50 (SD = 1.00) onthe correlation between patient appearance and agent path,and 0.25 (SD = 0.87) on the correlation between patient

<

PP-MoM

Figure 6. Mean rating differences between events testing cor-rectly matched and mismatched values of the target rule attributesin Experiment 2. SC = slate change; Env = environment; PA =patient appearance; AP — agent path; PP — patient path; MoM =agent manner of motion.

path and agent manner. As with rating accuracy, there was noevidence for an interaction, F(2, 30) < 1. The similaritybetween rating accuracy and interview scores was furtherhighlighted by a correlation of .87 (p < .001) between thetwo measures.

Discussion

Participants in Experiment 2 showed better learning of atarget rule when it formed part of a rich correlationalstructure than when no other correlations were present. Theresults of this experiment cannot be explained in terms ofcompetition among attributes or conflict among multiplepossible classifications for a participant's attention becauseonly one correlation was present in the condition in whichperformance was worse. The key difference between condi-tions thus seems to be value systematicity. Each target ruleattribute was predictive of the values of several otherattributes in the structured condition, whereas it was onlypredictive of one other attribute in the isolated condition.

In addition to the effects of correlational structure, bothExperiments 1 and 2 revealed differences in leamabilityamong the different target rules. Although it is difficult toaccount for these differences given the limited amount ofdata on event categories, the results of the next two experimentsoffer some suggestions as to what makes some correlations easierto learn than others. Further discussion of this issue follows thepresentation of the results of these experiments.

Experiment 3

Experiments 1 and 2 demonstrated facilitated learning ofevent correlations forming part of a rich correlational

648 KERSTEN AND BILLMAN

structure. Such correlational structure may be similar to thatassociated with instrument verbs. As we discussed previ-ously, however, the attributes associated with most verbscombine interchangeably with other event attributes, such asthose associated with prepositions or verb particles. Thus,unlike the events seen by participants in the structuredconditions of Experiments 1 and 2, most events are represen-tative of multiple, independent categories. For example, anevent involving running into a building can be thought ofeither as a running event or an into (i.e., entering) event.

We designed Experiment 3 to determine whether correla-tional structure facilitates the learning of event correlationsembedded within a matrix organization, similar to the onedescribed above. To this end, we used a mixed designinvolving one within-subjects variable and one between-subjects variable. We tested each participant on the learningof not one but rather two target rules constituting thewithin-subjects variable. One target rule formed part of asystem of correlations, whereas the other was isolated fromother correlations (see Figure 7). To control for rule diffi-culty, we counterbalanced the assignment of which rule wasstructured and which was isolated across two between-subjects conditions. Thus, for one group of participants, thepath rule formed part of a system of correlations and themanner rule was isolated, whereas for a second group, themanner rule was structured and the path rule was isolated.As a result, the learning of each target rule when isolatedcould be compared with learning when it was structured, justas in the previous experiments. We predicted that partici-pants would show facilitated learning of each target rulewhen it formed part of a system of correlations comparedwith when it was isolated.

We used two new target rules in this experiment toexplore further the generality of the effects of correlational

structure on event category learning. One was called thepath rule, involving a correlation between the paths of theagent and of the patient. For example, one category based onthis target rule involved events in which the patient ap-proached the agent until the two characters met, after whichthe agent pursued the patient in the other direction. Thesecond target rule was called the manner rule because itinvolved two aspects of the manner of motion of the agent,namely the motion of its legs relative to its body and theorientation of its body as it moved. For example, onecategory based on this target rule involved events in whichthe legs of the agent moved up and down along the length ofthe agent's body, causing the agent to move head first.

These two rules were involved in one of two configura-tions of correlations for each participant. For participants inthe structured path condition, two facilitator attributescovaried with the two path attributes. For participants in thestructured manner condition, the same two facilitatorscovaried instead with the two manner-of-motion attributes.We predicted better learning of each target rule when itformed part of a rich correlational structure than when it wasisolated, resulting in an interaction between the target rulebeing tested on a particular trial and the configuration ofcorrelations seen by a particular participant. Specifically, wepredicted better learning of the path rule in the structuredpath condition than in the structured manner condition, and wepredicted better learning of the manner rule in the structuredmanner condition than in the structured path condition.

Method

Participants

Thirty-six participants at the Georgia Institute of Technologyreceived course credit for their participation in this experiment.

Structured Path Condition

Leg motion # # Orientation

Stale change fl^—p»-^tt Environment

Hgent path 4 P ^ ^ ~ ^ ^ ^ J Patient path

Hgent appearance # • Patient appearance

Structured Hgent Condition

Leg motion Jtedat Orientation

State change • W L M H M H Z » Environment

flgent path Patient path

flgent appearance # # Patient appearance

Figure 7. Correlations seen by participants in Experiment 3.

Stimuli

All events. We made the events in Experiment 3 somewhatmore complex to accommodate the increased number of correla-tions associated with this design. Because the design required a setof four intercorrelated attributes together with an independent pairof correlated attributes, the existing seven attributes would haveleft only one of these attributes to vary randomly, perhaps makingthe task too easy. To make the task more difficult, we replaced theprevious attribute, manner of motion, with two attributes, resultingin a total of eight attributes in this experiment. One of these newattributes involved the orientation of the agent as it moved (seeFigure 8). Some agents moved in the directions they faced, somemoved sideways, and some backed up. To make this attributemeaningful, we changed the appearance of the agent from a squareto an elongated object with a head, tail, body, and legs. The secondnew attribute involved the leg motion of the agent, defined bydifferent motions of the legs relative to the body. Orientation andleg motion can together be considered to constitute an agent'smanner of motion because they implicate a particular mechanismfor achieving locomotion. We also made changes in the appearanceof the patient, making it more elongated with eyes at one end tomake it look similar to the agent. Finally, to preserve the speed ofthe animation despite this added complexity, we now defined theenvironment by pictures in unoccupied corners of the screen rather

EVENT CATEGORIES 649

Leg Motion

Orientat ion

00

;: H

Figure & Values of the two new motion attributes in Experiment3, along with two example values of the appearance of the agent.

than by background texture. The three environments were amountain, a swamp, and a desert.

Learning events. The learning events in Experiment 3 differedfrom those of the previous experiments in the correlations that werepresent Every participant was assigned the same two target rules:(a) the path rule, agent path-patient path; and (b) the manner rule,orientation-leg motion. In addition, state change and environmentcovaried with the two path attributes for participants in thestructured path condition, whereas these two attributes covariedinstead with the two manner-of-motion attributes for participants inthe structured manner condition.

Test events. To provide convergent validity for the finding offacilitation from correlational structure, we designed Experiment 3to involve a different test procedure from that of the previousexperiments. In particular, we used a forced-choice test rather thana rating task. In each test trial, we presented a participant with twoevents, one after the other. The participant's task was to choosewhich event was a better example of the events seen duringlearning. The two events varied on the value of one correlatedattribute. Half varied on agent path, testing for knowledge of theagent path-patient target rule, and half varied on orientation,testing for knowledge of the orientation-leg motion target rule. Weobscured the two facilitator attributes, state change and environ-ment, throughout testing in the same manner as in the previousexperiments, so that only knowledge of the target rule being testedcould be used to successfully choose the correct event. All otherattributes held constant values across the two events in each trial.There were 4 example trials, followed by 18 test trials, 9 of whichtested each target rule.

Design

There were two independent variables in Experiment 3. One wasthe target rule being tested (orientation-leg motion, agent path-patient path), manipulated within subjects. The other variable,configuration, was manipulated between subjects. This variablespecified which target rule participated in a system of correlationsand which was isolated (structured path or structured manner). Theprimary dependent variable was the number of correct choices inthe forced-choice test.

Procedure

The procedure of Experiment 3 differed from the previousexperiments in that participants were presented with only 90learning events to add further difficulty to the task. The testprocedure was also different. We instructed participants that theywere to choose which of two events was a better example of theevents seen during learning. They were first shown four exampletrials. In each trial, participants saw the first event, after which theypressed a button labeled Next Event to see the second event. Nochoices were required during the examples, with participantsinstead instructed after each pair of events about the different waysof obscuring attributes, as in die example test events of the previousexperiments. They were next presented with 18 test trials. Aftereach pair of events, participants pressed one of three buttons. Onebutton, labeled Repeat, allowed participants to see the two eventsagain. The other two buttons were labeled First Event and SecondEvent, allowing participants to indicate which event better exempli-fied the learning events.

Results

Figure 9 depicts the accuracy of participants in thestructured path and structured manner conditions on theforced-choice test, with 50% representing chance perfor-mance. Recall that the prediction of facilitation from correla-tional structure would receive support not from a main effectbut rather from an interaction of configuration and targetrule. An ANOVA on forced-choice accuracy revealed thisinteraction to be significant, F(l , 34) = 8.24, p < .01,MSE — 402.65. As we predicted, participants in the struc-tured path condition were more accurate on tests of the pathrule. These participants averaged 86% correct (SD = 20%),compared with an average of 68% (SD = 26%) in thestructured manner condition, t(34) = 2.34, p < .05 (one-tailed). On tests of the manner rule, participants in thestructured manner condition were more accurate. Theseparticipants averaged 61% (SD = 17%), compared with anaverage of 51% (SD = 13%) in the structured path condi-tion, f(34) = 1.81, p < .05 (one-tailed). Participants ingeneral performed much better with the path rule than withthe manner rule, producing a significant main effect of targetrule, F(l, 34) = 19.69, p < .001, MSE = 402.65. The main

oo

uu -

90

80-

70-

60-

50"//

T

//

• Structured path10 Structured manner

T

Path Rule Manner Rule

Figure 9. Mean rating accuracy on forced-choice trials testingknowledge of the two target rules in Experiment 3. The path ruleinvolved a correlation between agent path and patient path, and themanner rule involved a correlation between leg motion andorientation.

650 KERSTEN AND BILLMAN

effect of configuration did not approach significance,F (1 ,34 )<1 .

Interview scores revealed a similar pattern of results,although the interaction of configuration and target rule onlyapproached significance, F(l, 34) = 3.08, p < .09, MSE =1.30. With the path rule, participants in the structured pathcondition averaged 2.0 (SD = 1.28) correct reports of valuepairings, compared with an average of 1.28 (SD = 1.41) inthe structured manner condition. Although this differenceonly approached significance, f(34) = 1.61, p < .06(one-tailed), the trend in interview scores was quite similarto that in forced-choice accuracy with this rule, producing acorrelation of .77 (p < .01). Ten participants in the struc-tured path condition reported all three pairings of the path rule,compared with 6 participants in the structured manner condition.

With the agent rule, participants in the structured mannercondition averaged 0.33 (SD = 0.84) correct reports, com-pared with an average of 0.11 (SD = 0.47) in the structuredpath condition. Although this trend was in the same directionas the trend in forced-choice accuracy with this rule, thedifference between conditions was not significant, f(34) =0.98, p > .10. Moreover, the correlation between interviewscores and forced-choice accuracy was not significant forthis rule, r(34) = .32, p > .10. The failure of this correlationto reach significance is likely a result of a floor effect ininterview scores. Only 4 participants reported any knowl-edge of the manner rule: 3 participants in the structuredmanner condition and 1 participant in the structured pathcondition. Only 1 participant, in the structured mannercondition, reported all three pairings of the manner rule.Thus, there was much less reporting of the manner rule thanthe path rule, producing a significant main effect of targetrule, F(l9 34) = 27.69, p < .001, MSE = 1.30. The maineffect of configuration again did not approach significance,f (1, 34) = 1.16, p > .10, MSE = 0.97.

Leg motion seems to have been the source of difficulty forparticipants in reporting the manner rule. Although very fewparticipants reported associations between orientation andleg motion, 9 participants in the structured manner conditionreported associations between orientation and state change.Only 4 participants reported associations between leg mo-tion and state change, and 3 of these also reported associa-tions involving orientation. Thus, participants almost neverreported associations involving leg motion in the absence ofassociations between orientation and state change.

Discussion

The results of Experiment 3 revealed greater learning ofeach target rule when it formed part of a rich correlationalstructure than when it was isolated, even when another,unrelated correlation was present in the events. Participantsrevealed more knowledge of the path rule in the structuredpath condition than in the structured manner condition. Withthe manner rule, participants in the structured mannercondition performed better than participants in the structuredpath condition in the forced-choice test, whereas trends ininterview scores were not significant but in the predicteddirection. These results again provide evidence for facili-

tated learning of correlations involving high value systema-ticity. Independent correlations that were present in thestimuli allowed for the possibility of cross-classifying agiven event into more than one category, similar to the waymore than one verb can apply to a given event. Althoughverbs participating in such a matrix structure have beenargued to have little correlational structure (Huttenlocher &Lui, 1979), the existence of a matrix structure in the presentstimuli did not deter participants from discovering the richcorrelational structure associated with one of the target rules.

We designed Experiments 1-3 to determine whethercorrelational structure facilitates the learning of event catego-ries in a similar fashion to object categories. Apart from thisglobal issue of content, however, we have not put very muchfocus on the particular correlational rules used to instantiatethe structural relations of interest. Rather, we have tried touse a variety of event attributes across different experimentsto assess the generality of our findings. The pattern of resultshas been qualitatively very similar across all target rules, allshowing better learning in a structured context than whenisolated. Quantitatively, however, the rules have differedboth in their overall learnability and in the degree to whichthey showed facilitation from correlational structure.

It is likely that both learnability and degree of facilitationare a function of not only the salience of individual attributesbut also the relatedness of those attributes, in other words,the ability of participants to construct a hypothesis relatingthe attributes. In particular, the learnability of a rule maydepend on the relatedness of the attributes in that rule,whereas the degree of facilitation may depend on therelatedness of the rule attributes to the facilitator attributes.Although the general issue of event attribute relatedness istoo broad to be tackled in the present article, we next discusssome specific suggestions regarding the relatedness ofdifferent attributes found in the contrast between the twotarget rules in Experiment 3. After discussing the reasons forthis difference in learnability, we demonstrate in Experiment4 the effects of attribute relatedness on degree of facilitation,helping to clarify the conditions under which facilitationfrom correlational structure should be found.

Although we tried to select target rules of approximatelyequal difficulty, participants had much more difficulty learn-ing the manner rule, involving a correlation between legmotion and orientation. When this rule was isolated fromother correlations, participants performed no better thanchance, and participants in the structured manner conditionaveraged only slightly better than 60%. Interview datasuggest that leg motion was the source of difficulty forparticipants in learning this rule. This difficulty in learningassociations involving leg motion may stem from the factthat this attribute involved part motion, rather than themotion of an object as a whole, as was the case withorientation and the path attributes in this experiment andmanner of motion in the previous experiments. It is possiblethat attributes that are internal to a particular object, such asobject parts and their motions, are difficult to associate withmore global event attributes such as paths and state changes.People may be biased to link certain types of properties toobject categories and different properties to event categories.

EVENT CATEGORIES 651

Leg motion, although dynamic and analogous to mannerverbs in English (e.g., to walk, to strut), may have beenlinked more to object than to event categories in thisexperiment.

Orientation, in turn, seems to have played a mediatingrole between these two types of category. This makes sensegiven that orientation involves the motion of the agent as awhole, yet this motion is most easily defined in terms ofparts of the agent (e.g., direction of motion relative to thedirection the agent's head is pointing). Thus, people mayhave been able to infer the relation of leg motion to statechange only by first noticing the relation of state change toorientation and the relation of orientation to leg motion. Ifpeople indeed have difficulty associating object attributeswith global event attributes, this should have implicationsfor degree of facilitation. In particular, a covarying objectattribute should offer little facilitation to an event target rule,whereas a covarying event attribute should not facilitate anobject target rule. We investigated this prediction in Experi-ment 4.

participants who did not learn this correlation. In theobject-sorting task, we instructed participants to sort theagents appearing in the events. Participants who had knowl-edge of the object rule were expected to be more likely tobase these object sorts on agent body and agent legs thanparticipants who did not learn this correlation.

We thus made the same predictions for the rating task andthe sorting task. In particular, we predicted that participantswould show greater knowledge of each target rule when itformed part of a rich correlational structure than when it wasisolated. Moreover, we predicted greater facilitation of theobject rule when an object attribute (i.e., agent head) actedas facilitator than when a global event attribute (i.e., statechange) acted as facilitator. In contrast, we predicted greaterfacilitation of the event rule when a global event attributeacted as facilitator than when an object attribute acted asfacilitator. Thus, we predicted not only an interaction of theconfiguration of correlational structure with target rule, inreplication of Experiment 3, but also a three-way interactionof configuration, target rule, and facilitator attribute.

Experiment 4

In Experiment 4 we had three objectives. First, we soughtto replicate the findings of facilitation from correlationalstructure with the mixed design of Experiment 3. Second,we wanted to further investigate the notion that certainobject attributes are difficult to associate with global eventattributes. To do this, we compared the learning of two targetrules. The object rule was based on object attributes, namelythe body and legs of a complex agent. For example, onecategory based on this target rule involved agents withsquare, black bodies and three red legs on each side. Theevent rule was based on global event attributes, namelyagent path and environment. For example, one categorybased on this target rule involved events that took place on adesert background in which the agent pursued the patientafter contact. A third attribute covaried with the attributes inthe object rule for half of the participants and with theattributes in the event rule for the other half. This facilitatorattribute was either agent head, the object facilitator, or statechange, the event facilitator. If object parts are difficult toassociate with event properties, one would expect agenthead to facilitate the object rule to a greater extent than statechange, even though state change may be much moresalient. In contrast, the event rule would be expected to showmore facilitation from state change than from agent head.

A third objective of Experiment 4 was to study the effectsof correlational structure on a more traditional measure ofcategory learning, the sorting of instances into categories(Fried & Holyoak, 1984; Homa & Cultice, 1984). We alsowanted to examine the relationship of this measure to therating measure used in the previous experiments. After therating task, we administered two different sorting tasks, eachtesting for knowledge of one of the two target rules. In theevent-sorting task, we instructed participants to sort entireevents into three categories. Participants who learned theevent rule were expected to be more likely to base theirevent sorts on agent path and environment than were

Method

Participants

Thirty-two undergraduates at the Georgia Institute of Technol-ogy received course credit for their participation in this experiment.

Stimuli

All events. The events in Experiment 4 differed from those inExperiment 3 in that the head, body, tail, and legs of the agentvaried independently of one another, rather than always appearingin the same three combinations for a given participant. Orientationno longer varied in this experiment. Instead, all characters movedhead first.

Learning events. The learning events in Experiment 4 differedfrom those in Experiment 3 in the correlations that were present.We assigned every participant the same two target rules: (a) theevent rule, agent path-environment; and (b) the object rule, agentbody-agent legs. In addition, a third attribute covaried with theevent rule for participants in the structured event condition or withthe object rule for participants in the structured object condition.State change played this role for half of the participants, whereasagent head did so for the other half.

Test events. In Experiment 4 we employed the rating task usedin Experiments 1 and 2 rather than the forced-choice task used inExperiment 3. The forced-choice task would likely have influencedparticipants' sorting performance because only one attribute variedacross events in each forced-choice trial, making participants morelikely to realize the importance of that attribute and thus use it asthe basis for their sorts. Participants rated 36 test events, with 18testing each target rule. The facilitator attribute for a particularparticipant was obscured throughout testing. In addition, eitherenvironment or agent body was obscured in each event, dependingon which target rule was being tested. Body parts were obscured byblackening them, similar to the way the agent and patient as awhole were blackened in the previous experiments. State changeand environment were obscured as in the previous experiments.

Sorting trials. After the test events, we presented each partici-pant with 36 sorting trials, divided into two groups of 18. For onegroup, we instructed the participant to sort the events into three

652 KERSTEN AND BILLMAN

categories. After each event, three buttons appeared at the bottomof the screen. Initially, these buttons were labeled Category 1,Category 2, and Category 3, respectively. Participants were able tochange these to more meaningful labels, however, by clicking oneach label and typing a different label. For the other group ofevents, we instructed the participant to sort the agents appearing inthe events into three categories. The procedure was otherwise thesame.

Throughout the sorting trials, the facilitator attribute was ob-scured. This was done so that participants would not sort on thefacilitator attribute. Otherwise, participants would be more likelymerely by chance to be credited with a sort consistent with theattributes participating in a richer correlational structure. Obscur-ing the facilitator attribute equated the baseline probability ofchoosing an attribute from either target rule as the basis for sorting.

Design

The primary dependent variable in Experiment 4 was thedifference between each participant's average rating for eventsinvolving correctly matched values of the target rule attributes andhis or her average rating for events involving mismatched values.There were three independent variables. One was the target rulebeing tested (event rule or object rule), manipulated withinsubjects. The other two variables were configuration, specifyingwhich target rule formed part of a rich correlational structure andwhich was isolated (structured event or structured object), and theidentity of the facilitator attribute (event facilitator or objectfacilitator), both manipulated between subjects.

Procedure

We presented 90 learning events by using the same procedure asin Experiment 3. Thirty-six test events then followed, with thesame rating task as in Experiments 1 and 2. The rating task wasfollowed by 36 sorting events. We instructed participants to sort 18events on the basis of the agents in those events. They were toldthat there were three kinds of agents (referred to as "Truzioids"throughout the experiment) on this planet, and that they were toindicate which type of agent they saw in each event by clicking onone of the three buttons at the bottom of the screen. Participantswere allowed to change the labels on the buttons as soon as theyhad finished the instructions for the sorting task, but they were notobligated to do so and could change a given label more than once.The instructions for the event-sorting task were the same exceptthat we instructed participants to sort events instead of agents. Halfof the participants sorted events first, and half sorted agents first.The labels on the buttons reverted to Category I, Category 2, andCategory 3 before beginning the second sorting task.

Table 6Rating Accuracy With Event Rule in Experiment 4

Results

Rating Data

Tables 6 and 7 display the mean ratings of events testingthe two target rules for participants in the structured eventand structured object conditions, and Figure 10 depictsrating differences. An ANOVA revealed no significant maineffects of configuration, F(U 28) = 2.50, p > .10, MSE =0.99, facilitator attribute, F(l, 28) = 1.48, p > .10, MSE =0.99, or target rule, F(l, 28) = 2.75, p > .10, MSE = 0.99.The prediction of facilitation from correlational structure,however, would receive support not from a main effect but

Condition

Structured eventEvent facilitatorObject facilitator

Structured objectEvent facilitatorObject facilitator

Incorrectevents

M

2.321.762.873.683.653.72

SD

0.990.840.840.760.550.96

Correctevents

M

3.674.113.233.813.683.94

SD

0.890.760.830.780.860.73

Difference

M

1.352.350.360.120.030.22

SD

1.731.511.360.670.440.86

Note. The event rule involved a correlation between agent pathand environment. The event facilitator was state change, and theobject facilitator was agent head.

rather from an interaction of the configuration of correla-tions with target rule. This interaction of target rule andconfiguration was significant, F (1, 28) = 12.30, p < .01,MSE = 0.91, replicating Experiment 3. Rating accuracy washigher with each target rule when it formed part of a richcorrelational structure. With the event rule, rating accuracywas higher in the structured event condition (M ^ 1.35,SD = 1.73) than in the structured object condition (M = 0.12,SD = 0.67), f(30) = 2.65, p < .01 (one-tailed). With theobject rule, rating accuracy was higher in the structuredobject condition (M = 0.57, SD = 0.83) than in the struc-tured event condition (M - 0.12, SD - 0.73), although thisdifference only approached significance, r(30) = 1.61, p <.06 (one-tailed).

As we predicted, the pattern of facilitation we describedabove was moderated by the choice of facilitator attribute,resulting in a significant three-way interaction of target rule,configuration, and facilitator, F(l , 28) = 4.20, p < .05,MSE = 0.91. With the event rule, the advantage of thestructured event condition over the structured object condi-tion was for the most part carried by participants who wereassigned the event facilitator (i.e., state change). A post hocFisher's least significant difference (LSD) test revealed thatrating accuracy with the event rule was higher in thiscondition than in any of the other three combinations ofconfiguration and facilitator (p < .05), whereas there wereno differences among these other three groups. With the

Table 7Rating Accuracy With Object Rule in Experiment 4

Condition

Structured eventEvent facilitatorObject facilitator

Structured objectEvent facilitatorObject facilitator

Incorrectevents

M

3.313.523.093.323.263.38

SD

0.900.900.900.620.670.60

Correctevents

M

3.433.553.303.893.624.15

SD

0.490.580.380.700.580.74

Difference

M

0.120.030.210.570.360.77

SD

0.730.490.940.830.820.84

Note. The object rule involved a correlation between agent bodyand agent legs. The event facilitator was state change, and theobject facilitator was agent head.

EVENT CATEGORIES 653

Event Facilitator

c

3

2

1 •

0

TStructured eventStructured object

• •

Event Rule Object Rule

rac

yA

cc

iti

ng

CE

V

2 -

1 -

n -

Object

•CD

Facilitator

Structured eventStructured object

T -

Event Rule Object Rule

Figure 10. Mean rating differences between events testing cor-rectly matched and mismatched values of the target rule attributesin Experiment 4. The path rule involved a correlation betweenagent path and environment, whereas the agent rule involved acorrelation between agent body and agent legs. The top paneldepicts performance when state change was correlated with one ofthese two pairs of attributes, whereas the bottom panel depictsperformance when agent head acted as facilitator.

object rule, the advantage of the structured object conditionwas greater with the object facilitator (i.e., agent head) thanwith the event facilitator, although an LSD test showed nosignificant differences between groups.

This pattern of results also produced two other significantinteractions. First, there was an interaction of target rule andfacilitator, F(l , 28) = 6.17, p < .05, MSE = 0.91. Ratingaccuracy on the event rule was higher for participants whowere assigned the event facilitator, whereas accuracy on theobject rule was higher for participants assigned the objectfacilitator. Second, there was a significant interaction ofconfiguration and facilitator, F(l, 28) = 5.85, p < .05,MSE = 0.99. Aggregating the two rules, participants in thestructured event condition performed more accurately withthe event facilitator, whereas participants in the structuredobject condition performed better with the object facilitator.

Interview Data

Interview scores also revealed an interaction of configurartion and target rule, F(l, 28) = 5.45,p< .05, MSE = 0.83.The event rule was reported more often in the structuredevent condition (M = 1.25, SD = 1:44) than in the struc-tured object condition (M = 0.06, SD - 0.25), f(30) = 3.26,p < .01 (one-tailed). Six participants in the structured eventcondition reported all three pairings of the values of theevent rule attributes, but none did in the structured objectcondition. The two groups did not differ in reports of theagent rule, however, with means of 0.38 {SD — 0.89) in the

structured event condition and 0.25 (SD = 0.68) in thestructured object condition, ((30) = 0.45, p > .10. Only 1participant, in the structured event condition, reported allthree pairings with this rule.

As with rating accuracy, this pattern of facilitation withinterview scores was moderated by the choice of facilitatorattribute, producing a three-way interaction, F(l , 28) =4.25, p < .05, MSE = 0.83. With the event rule, participantsassigned to die structured event condition with the eventfacilitator reported more knowledge than did participants inthe structured object condition assigned either facilitator(p < .05). There were no other significant differencesbetween groups, nor were there any significant differencesbetween groups on reports of the agent rule.

The only other significant effect was a main effect ofconfiguration, F(l , 28) = 8.70, p < .01, MSE = 0.79, withparticipants in the structured event condition performingbetter than participants in the structured object condition.Overall, the trends in interview scores and rating accuracywere quite similar for the event rule, as indicated by thecorrelation between these two measures, r = .71,p < .01. Incontrast, there was no significant relation between the twomeasures with the object rule, r = .01, p > .10.

Sorting Data

We generated two sorting scores. The event sorting scorereflected the degree to which participants sorted events onthe basis of the attributes in the path rule, agent path andenvironment. The object sorting score reflected the degree towhich participants sorted agents on the basis of the attributesin the object rule, agent body and legs.

Sorting scores indexed the number of changes that wouldhave to be made in a participant's sort to make it entirelyconsistent with the values of the target rule attributes. Tocompute this score, we first determined which of the threecategories was assigned most consistently to a single pair ofvalues of the target rule attributes. We then counted thenumber of times this category was assigned to other valuesof the target rule attributes. Next, we determined which ofthe remaining two categories was assigned more consis-tently to one of the remaining two pairs of values. Thenumber of times this category was assigned to the other pairof values was then added to the previous count. We finallyadded to the sorting score the number of times the thirdcategory was assigned to a pair of values other than theremaining pair. Thus, a score of 0 reflected a sort that wasperfectly consistent with the values of the target ruleattributes, and the highest possible score was 12 (with 18 ofeach type of sorting event, the participant could not help butget at least 6 right). We predicted that event sorting scoreswould be lower (and thus better) in the structured eventcondition than in the structured object condition, whereasobject sorting scores would be lower in the structured objectcondition.

We performed an ANOVA on sorting scores with configu-ration and facilitator as between-subjects variables andtarget rule (event sort vs. object sort) as a within-subjectvariable. Preliminary analyses revealed no significant main

654 KERSTEN AND BILLMAN

effect or interactions involving the order of the two tasks, soorder was not included as an independent variable in thefinal analysis. Consistent with rating and interview data, thisanalysis revealed an interaction of configuration and targetrule, F(l, 28) = 4.97,/? < .05, MSE = 12.87. Event sortingscores were lower in the structured event condition(M = 2.75, SD - 3.19) than in the structured object condi-tion (M = 5.00, SD = 4.05), f(30) = 1.75, p < .05 (one-tailed). Seven participants in the structured event conditionproduced sorts that were entirely consistent with values ofthe event rule attributes, compared with 4 in the structuredobject condition. Object sorting scores, in turn, were lowerin the structured object condition (M = 4.81, SD = 5.12)than in the structured event condition (Af = 6.56, SD = 4.75),although this difference was not significant, f(30) = 1.00,p > .10. Five participants in the structured object conditionproduced sorts that were entirely consistent with the valuesof the object rule attributes, compared with 4 in thestructured event condition.

The only other effect to attain significance was the maineffect of facilitator, F(U 28) = 5.92, p < .05, MSE = 22.33.Sorting scores were lower with the event facilitator, averag-ing 2.31 (SD = 3.22) and 4.38 (SD = 4.86) on the eventsorting score and object sorting score, respectively, com-pared with averages of 5.44 (SD ~ 3.71) and 7.00(SD = 4.80) on the same two sorts with the object facilitator.The main effect of target rule also approached significance,f ( l , 28) = 4.08, p < .06, MSE = 12.87, with lower scoreson event sorts (Af = 3.88, SD = 3.77) than on object sorts(M = 5.69, SD = 4.93).

As we predicted, the trend in event sorting scores washighly related to the trends with the event rule in both ratingaccuracy (r — — .50, p < .01) and interview scores(r = —.56, p < .01), with the correlations being negativebecause lower sorting scores reflected better performance. Incontrast, object sorting scores were not significantly relatedto rating accuracy (r = .25, p > .10) or interview scores(r— - . 12, p > . 10) with the object rule.

Discussion

Experiment 4 revealed that an object attribute facilitatedthe learning of an object correlation to a greater extent thandid an event attribute, whereas the event attribute producedmore facilitation of an event correlation. A correlationbetween two body parts, agent body and agent legs, showedlittle facilitation from a covarying state change, even thoughstate change was found to be a highly salient facilitator ofevent correlations in this and earlier experiments. The objectrule was facilitated to a greater extent by a covarying objectpart, namely agent head. This finding is consistent with otherwork showing facilitated learning of object correlations inthe presence of additional, covarying object properties(Billman & Knutson, 1996). In contrast, the event ruleshowed much greater facilitation from state change thanfrom agent head. Taken together, these findings provideevidence for a certain degree of encapsulation of objectproperties within the representations of events in whichthose objects participate. More generally, the degree of

facilitation from correlational structure is a function not onlyof the saliences of the individual attributes involved but alsothe relatedness of those attributes.

Sorting scores as well as rating scores replicated theinteraction of the configuration of correlational structurewith target rule found with the rating task in Experiment 3.Participants were more likely to sort on the basis of thetarget rule attributes when a third attribute covaried withthose attributes. In particular, participants were more likelyto sort events on the basis of the attributes of the event rule inthe structured event condition than in the structured objectcondition, whereas they were more likely to sort agents onthe basis of the attributes of the object rule in the structuredobject condition. This finding of convergence between ameasure of correlation learning and a more traditionalmeasure of category learning provides evidence that thelearning of event correlations can be taken as a measure ofevent category learning.

Event sorting scores were significantly correlated withrating accuracy on the event rule, indicating that participantswho learned the agent path-environment correlation tendedto sort events on the basis of the values of one or both ofthese attributes. Although the correlation between thesemeasures was not overwhelming, it is likely to have beenattenuated to some extent by participants who failed to learnany correlations but by chance happened to choose one ofthe correlated attributes as the basis for sorting. Thisoutcome was particularly likely with participants for whomstate change was obscured during sorting. These participantshad to choose one of the other attributes as a basis forsorting, with the most "eventlike" of the remaining at-tributes being agent path, environment, and patient path, twoof which participated in the event rule.

Unlike event sorts, object sorting scores were not signifi-cantly correlated with rating accuracy on the object rule.This null result may be related to the lack of correlationbetween rating accuracy and interview scores with thistarget rule. Some participants were apparently tappingimplicit knowledge of this target rule when performing therating task. A lack of explicit knowledge of the agentbody-agent legs correlation may explain why these partici-pants were no more likely than other participants to basetheir object sorts on these attributes. Particularly compellingevidence for an influence of implicit knowledge on ratingaccuracy comes from participants in the structured objectcondition assigned the object facilitator. These participantsrated correct events testing the object rule more than fa of apoint higher than they did incorrect events, but not one ofthese participants reported any knowledge of this correlationin postexperimental interviews.

It thus seems that rating performance may have beeninfluenced by partial, difficult-to-articulate knowledge towhich the interview and sorting tasks were not sensitive. Asa result, rating data from Experiment 4 provide evidence forfacilitated learning of each target rule when it formed part ofa rich correlational structure, whereas interview and sortingscores only reveal facilitation for the target rule of whichparticipants had explicit knowledge, namely the event rule.

EVENT CATEGORIES 655

General Discussion

In the research reported here we examined the unsuper-vised learning of event categories based on differentlyorganized sets of correlations among a number of attributes.The results of four experiments revealed facilitated learningof correlations when they formed part of a rich correlationalstructure. Of the 10 target rules we used in these fourexperiments, all 10 showed the predicted pattern. Experi-ment 1 revealed that an individual correlation was learnedbetter when forming part of a rich correlational structurethan when independent of other correlations. Experiment 2showed that this effect could not be explained entirely interms of competition among independent correlations in thematrix condition, revealing the same effect when no indepen-dent correlations were present. Experiment 3 revealedfacilitated learning of correlations forming part of a richcorrelational structure, even when that structure was embed-ded within a matrix organization. Experiment 4 providedconverging evidence for facilitated learning of categoriesbased on rich correlational structure, revealing effects ofcorrelational structure on the sorting of events into catego-ries. Experiment 4 also provided evidence regarding the roleof object categories in event representations, as we discussbelow.

The Relation Between Object and Event Categories

The pattern of results found in these experiments issimilar to that found with experiments on the unsupervisedlearning of object categories (Billman & Knutson, 1996).Apparently, similar learning biases operate in both domainsto facilitate the learning of correlations forming part of a richcorrelational structure. Coupled with the observation thatverb meanings have much less correlational structure thando nouns, however, this rinding poses a puzzle. It is possiblethat such structure is indeed rare in the domain of events,accounting for the weaker correlational structure associatedwith verbs. But an alternative is that events do have a richcorrelational structure, which people know about, but factorsspecifically linked to language are responsible for the lack ofcorrelational structure in verb meanings. That is, the role ofverbs may be to select one aspect of an event as being mostimportant to its description, with additional detail specifiedby its arguments.

Verbs may therefore express a relatively abstract level ofevent categories. The predictive work carried by verbs mayprimarily emerge in combination with its arguments. Forexample, the meaning of the verb to chase in isolation seemsto primarily convey the notion of one character followingafter another in the same direction, perhaps along withimplications for speed and effort on the part of bothparticipants in the event. When the objects cheetah andgazelle are inserted into these argument slots, however,many more inferences are possible, including a possibleoutcome. Thus, object and event categories may be funda-mentally similar in structure and in the biases that affect theirlearning, but additional factors may constrain the structureof verb meanings.

Learnability may be one factor that affects the structure ofverb meanings. Talmy (1985) has suggested that an immenselexicon would be required if a language involved manyverbs that expressed more than a few attributes of meaning.For example, a language involving verbs with specificmeanings such as "cheetah chases gazelle" would require averb for each combination of objects that take part in eventsinvolving chasing. The large number of such verbs coupledwith the low frequency of use of each individual verb wouldmake such a language very difficult to learn.

If object and event categories are indeed fundamentallysimilar in the way they are learned and in their resultingmental representations, how do these two types of categorydiffer? It is possible that they reflect two different levels in arepresentational hierarchy. The common motion of the partsof an object may segregate that object from the rest of anevent, producing an encapsulated representation for thatobject (Kellman & Spelke, 1983). This encapsulation wouldexplain the results of Experiments 3 and 4, which revealedthat participants were much better at associating object partsand their motions with other object-related attributes thanwith more global event attributes. The role of objectinformation in event categories may thus be mediated byobject category representations. Event representations mayinclude labels or pointers to the categories of objects playingroles within those events, but obtaining further objectinformation may require consulting object category represen-tations directly.

This hypothesis allows one to integrate the results ofExperiment 4 with earlier work on the perception ofcausality. In particular, Cohen and Oakes (1993) found that10-month-old infants associated agents in causal events withtheir effects on other objects, as indicated by dishabituationwhen an agent produced the wrong effect. This is similar toMichotte's (1946/1963) finding that the effect in a causalevent was perceived as belonging to the agent, even whenonly the patient carried the effect. Both of these findings,however, involve a relation between an effect and a wholeobject, rather than an object part. It is possible that peoplecategorize objects independently of their effects on otherobjects, on the basis of object features such as parts and partmotions. This conclusion is suggested by the finding ofExperiment 4 that state change produced little facilitation ofthe correlation between agent body and agent legs, appar-ently because it was difficult for participants to relate statechange to object parts. The resulting object categories,however, may become associated with characteristic effectsof that category of objects on other objects, representing thisinformation as part of an event category.

There thus seems to be a part-whole relation betweenobjects and events, with objects forming part of an eventcategory representation, if only at the level of a categorylabel or pointer. One could argue, in fact, that objects play asimilar role in event representations as do features in objectrepresentations, if one believed object features to be catego-ries in their own right rather than primitives (Schyns,Goldstone, & Thibaut, in press). Clearly, however, eventsare composed of other types of features besides objects. Forexample, the role of temporal information in events would

656 KERSTEN AND BILLMAN

be difficult to account for solely in terms of object informa-tion. Although we have suggested that object categoriesinclude information about the characteristic motions ofthose objects (see also Kersten & Billman, 1995) and thatrepresenting such motions may require reference to time,such motions seem to be continuous and repeating. Incontrast, events often seem to have a much more complextemporal structure, often involving long durations ending inan irreversible outcome. Clearly, understanding the role oftime in representations is central to understanding differ-ences between object and event categories.

Implications for Category Learning Models

Many more models have been proposed to account forsupervised learning than for unsupervised, and most catego-rization experiments tend to focus exclusively on supervisedlearning. Although these models were not designed to beapplied to unsupervised learning, adapting them to modelthe present task could be useful in trying to understanddifferences between supervised and unsupervised learning.One possibility would be to treat one of the target ruleattributes as analogous to a category label. Thus, the task fora model would be to learn to predict the values of this targetrule attribute on the basis of the values of other attributes.

When adapted in this way, prominent models of categorylearning (e.g., Gluck & Bower, 1988; Kruschke, 1992;Trabasso & Bower, 1968) predict effects opposite to thosefound in our experiments. In these models, multiple predic-tors of an attribute compete for predictive strength. Becausepredictive strength is treated as limited, learning the predic-tiveness of one cue reduces the likelihood of learning thepredictiveness of a second cue. In the present task, therewere multiple predictors of each attribute that participated ina system of correlations. According to competitive learningmodels, these predictors should compete for predictivestrength, and thus, correlations should be better learnedwhen isolated than when forming part of a system ofcorrelations. For example, in Experiment 4, participantsshould have been less likely to learn that environment waspredictive of agent path if state change was also predictive.

The fact that supervised learning models cannot accountfor the present results does not invalidate these as models ofsupervised learning so much as to point to differencesbetween supervised and unsupervised learning. It is notsurprising that cue competition occurs in supervised tasksbecause it is clear to participants that the prediction ofcategory membership is sufficient for successful perfor-mance in these tasks, and thus they need look no further oncethey have discovered an adequate predictor. In contrast, aparticipant may be more open to new information when it isnot as clear what information is required for the task.

Differences in temporal relations, in and of themselves, donot seem to be responsible for the different results found insupervised and unsupervised learning tasks. The temporalrelations in the present unsupervised task were quite similarto those in many supervised tasks, with some types ofinformation revealed before others. A few participants didreport treating state change as the attribute to be predicted in

this task, similar to a category label in supervised learningtasks. These participants generally demonstrated knowledgeof only one predictor of state change, similar to findingsfrom supervised tasks. As indicated by the group results,however, most participants did not treat the task in this way.The key difference between supervised and unsupervisedtasks thus seems to be the singling out of one type ofinformation as most important to predict. A useful avenue offuture research may be to assess how well different categori-zation tasks in the world map onto supervised and unsuper-vised learning in the laboratory.

We next consider one model of unsupervised learning(Billman & Heit, 1988) that accounts for the results of theseexperiments. Relatively few models have been designedspecifically to accpunt for unsupervised learning (Anderson,1991; Billman & Heit, 1988; Martin, 1992; Schyns, 1991),though a number have been developed in the machinelearning community (e.g., Fisher, 1987). Nor have therebeen many applications of supervised models to conceptlearning tasks that lack discriminative feedback (Heit, 1994;Medin, Altom, Edelson, & Freko, 1982). Some of theseother models might be able to account for our results byextending their implementations to address our task, and wewould be interested in the results of such extensions.

The internal feedback model of unsupervised learning(Billman & Heit, 1988) predicts the pattern of faciliation wesought and found in the current experiments. Unlike super-vised learning models, this model does not single out aparticular attribute as being most important to predict.Instead, the model chooses in each trial a particular attributewhose value is to be predicted. This prediction is made onthe basis of an internalized rule involving the value of asecond, predictor attribute. The predicted value of thisattribute is then compared with the actual, presented value ofthe attribute, providing internal feedback as to the predictivepower of the rule.

Initially, all rules are equal in strength, and thus thesystem is no more likely to predict one value of an attributeover another. Yet with predictive success, a rule is strength-ened and as a result is used more often to predict the correctvalue of the attribute in question. Critically, however, eachrule is specific to a particular pair of attributes. Rules areselected only if the relevant attributes are sampled, andattributes are sampled proportional to their salience. Fo-cused sampling is the mechanism that allows the model toaccount for benefit from rich correlational structure. When acorrect prediction is made, the salience of the attributesinvolved increases. As a result, these attributes are sampledmore frequently and thus other rules in which those at-tributes play a part become more likely to be learned. Theinternal feedback model thus predicts that a correlation willbe better learned when it forms part of a system ofcorrelations than when it is in isolation.

We illustrate the focused sampling procedure by describ-ing the actions of a hypothetical participant in the structuredevent condition in Experiment 4. On discovering a correla-tion between state change and environment, the participantwould allocate extra attention to those attributes, takingattention away from each attribute not involved in that

EVENT CATEGORIES 657

correlation. Because attention would be diverted away froma number of attributes toward only two attributes, however,less attention would be taken away from each attribute notyet discovered to be correlated than would be redirectedtoward each correlated attribute. As a result, each pairwisecombination of state change, environment, and agent pathwould receive more total attention than if no attentionalreallocation had occurred, making the participant morelikely to discover the target correlation between environ-ment and agent path.

Focused sampling is a useful procedure for learningcategories formed around rich correlational structure be-cause the attributes participating in such structure arepredictive of many other attributes. Thus, allocating atten-tion to attributes found to be predictive makes finding othercorrelations that form those categories more likely. Such anattentional reweighting procedure could be added to othermodels to make them more consistent with the presentfindings, as well as previous findings of facilitation fromcorrelational structure (Billman, 1989; Billman & Knutson,1996; Cabrera & Billman, 1996).

Interactions of Content With Correlational Structure

As we noted earlier, the different correlational rules weused in these experiments varied not only in generallearnability but also in degree of facilitation from correla-tional structure. The internal feedback model may help toclarify how these two issues are related. According to theinternal feedback model, facilitation of a particular targetrule occurs when one first notices a correlation between oneof the two target rule attributes and a third, facilitatorattribute. The attentional reallocation resulting from thisdiscovery encourages the discovery of other correlationsinvolving these attributes, such as the target rule. Thus,facilitation is most likely to be found when a correlationinvolving the facilitator attribute is easier to learn than thetarget rule itself. If the target rule were much easier to learnthan the correlations involving the facilitator attribute, onewould expect little facilitation because the correlationsinvolving the facilitator would not be learned until after thetarget rule and thus would not aid in its discovery. Inaddition, if a target were very difficult to learn, additionalattention directed to that rule may still not be sufficient forone to notice that correlation. Degree of facilitation may thusbe a function of the learnability of the target rule relative tothe other correlations present.

There remain many open questions as to the factorsinfluencing the learnability of correlational rules. On thebasis of the results of Experiments 3 and 4, it seems likelythat a theory of learnability will have to account for attributerelatedness as well as the individual salience of attributes.People may never hypothesize a relation between certainattributes even if they are individually very salient. Forexample, participants in Experiment 3 apparently had diffi-culty associating leg motion with state change, even thoughboth were subjectively very salient. In contrast, people mayoverestimate the degree of correlation between attributesthat intuitively seem like they should be related. For

example, Chapman and Chapman (1967) found that bothclinicians and naive judges overestimated the frequency ofco-occurrence of the trait "suspiciousness" with unusualdrawings of eyes in the Draw-a-Person test, apparentlybecause suspiciousness is intuitively associated with theeyes. Noticing a correlation between attribute values in anevent may depend on the availability of past instances inwhich those values were paired (Tversky & Kahneman,1973) or on the ability of learners to construct a theory as tohow those values may be related (Murphy & Medin, 1985).

Conclusions

In summary, these experiments make several distinctcontributions. First, the results provide evidence that peoplehave biases that facilitate the learning of richly structuredcategories, in the domain of events as well as of objects.Second, these experiments expand our knowledge of learn-ing from observation when supervision or feedback are notprovided. Finally, this research provides a useful buildingblock in the study of event categories, encouraging furtherstudy of why some types of event correlations are easier tolearn than others.

References

Anderson, J. R. (1991). The adaptive nature of human categoriza-tion. Psychological Review, 98, 409-429.

Barsalou, L. W., & Billman, D. (1988). Systematicity and semanticambiguity. In D. S. Gorfein (Ed.)> Resolving semantic ambiguity(pp. 146-203). New York: Springer-Verlag.

Behrend, D. A. (1990). The development of verb concepts:Children's use of verbs to label familiar and novel events. ChildDevelopment, 61, 681-696.

Billman, D. (1989). Systems of correlations in rule and categorylearning: Use of structured input in learning syntactic categories.Language and Cognitive Processes, 4(2), 127-155.

Billman, D., & Heit, E. (1988). Observational learning frominternal feedback: A simulation of an adaptive learning method.Cognitive Science, 12, 587-625.

Billman, D., & Knutson, J. F. (1996). Unsupervised conceptlearning and value systematicity: A complex whole aids learningthe parts. Journal of Experimental Psychology: Learning,Memory, and Cognition, 22, 458-475.

Cabrera, A., & Billman, D. (1996). Language-driven conceptlearning: Deciphering "Jabberwocky." Journal of ExperimentalPsychology: Learning, Memory, and Cognition, 22, 539-555.

Chapman, L. J., & Chapman, J. P. (1967). Genesis of popular buterroneous psycho-diagnostic observations. Journal of AbnormalPsychology, 72, 193-204.

Cohen, L. B., & Oakes, L. M. (1993). How infants perceive asimple causal event. Developmental Psychology, 29,421-433.

Fisher, D. (1987). Knowledge acquisition via incremental concep-tual clustering. Machine Learning, 2, 139-172.

Fried, L. S., & Holyoak, K. J. (1984). Induction of categorydistributions: A framework for classification learning. Journal ofExperimental Psychology: Learning, Memory, and Cognition,10, 234-257.

Gentner, D. (1981). Some interesting differences between verbsand nouns. Cognition and Brain Theory, 4,161-178.

Gluck, M. A., & Bower, G. H. (1988). From conditioning tocategory learning: An adaptive network model. Journal ofExperimental Psychology: General, 117, 227-247.

658 KERSTEN AND BILLMAN

Graesser, A. C , Hopkinson, P. L., & Schmid, C. (1987). Differ-ences in interconcept organization between nouns and verbs.Journal of Memory and Language, 26, 242-253.

Heit, E. (1994). Models of the effects of prior knowledge oncategory learning. Journal of Experimental Psychology: Learn-ing, Memory, and Cognition, 20, 1264-1282.

Homa, D., & Cultice, J. C. (1984). Role of feedback, category size,and stimulus distortion on the acquisition and utilization ofill-defined categories. Journal of Experimental Psychology:Learning, Memory, and Cognition, 10, 83-94.

Huttenlocher, J., & Lui, F. (1979). The semantic organization ofsome simple nouns and verbs. Journal of Verbal Learning andVerbal Behavior, 18, 141-162.

Huttenlocher, J., Smiley, P., & Chamey, R. (1983). Emergence ofaction categories in the child: Evidence from verb meanings.Psychological Review, 90, 72-93.

Jackendoff, R. (1987). On beyond zebra: The relation of linguisticand visual information. Cognition, 26, 115-122.

Kellman, P. J., & Spelke, E. S. (1983). Perception of partlyoccluded objects in infancy. Cognitive Psychology, 15,483-524.

Kersten, A. W., & Billman, D. (1995). TTie roles of motion andmoving parts in noun and verb meanings. In J. D. Moore & J. FLehman (Eds.), Proceedings of the Seventeenth Annual Confer-ence of the Cognitive Science Society (pp. 625-630). Hillsdale,NJ: Erlbaum.

Kruschke, J. K. (1992). ALCOVE: An exemplar-based connection-ist model of category learning. Psychological Review, 99, 22-44.

Leslie, A. M. (1988). The necessity of illusion: Perception andthought in infancy. In L. Weiskrantz (Ed.), Thought withoutlanguage (pp. 185-210). Oxford, England: Clarendon Press.

Leslie, A. M.» & Keeble, S. (1987). Do six-month-old infantsperceive causality? Cognition, 25, 265-288.

Malt B. C , & Smith, E. E. (1984). Correlated properties in naturalcategories. Journal of Verbal Learning and Verbal Behavior, 23,250-269.

Martin, J. D. (1992). Direct and indirect transfer: Investigations inconcept formation (Tech. Rep. No. GrTAI-92-58). Atlanta, GA:Georgia Institute of Technology, Department of ComputerScience.

Medin, D. L., Altom, M. W., Edelson, S. M., & Freko, D. (1982).

Correlated symptoms and simulated medical classification. Jour-nal of Experimental Psychology: Learning, Memory, and Cogni-tion, 8, 37-50.

Michotte, A. (1963). The perception of causality (T. R. Miles &Elaine Miles, Trans.). New York; Basic Books. (Original workpublished 1946)

Murphy, G. L.t & Medin, D. L. (1985). The role of theories inconceptual coherence. Psychological Review, 92, 289-316.

Nelson, K. (1986). Event knowledge: Structure and function indevelopment. Hillsdale, NJ: Erlbaum.

Rosch, E. H., Mervis, C. B., Gray, W. D., Johnson, D. M., &Boyes-Braem, P. (1976). Basic objects in natural categories.Cognitive Psychology, 8, 382-439.

Schank, R. C , & Abelson, R. P. (1977). Scripts, plans, goals, andunderstanding. Hillsdale, NJ: Erlbaum.

Schyns, P. G. (1991). A modular network model of conceptlearning. Cognitive Science, 15, 461-508.

Schyns, P. G., Goldstone, R. L., & Thibaut, J. (in press). Thedevelopment of features in object concepts. Behavioral andBrain Sciences.

Slobin, D. I. (1981). The origins of grammatical encoding ofevents. In W. Deutsch (Ed.), The child's construction of language(pp. 185-199). London: Academic Press.

Talmy, L. (1985). Lexicalization patterns: Semantic structure inlexical forms. In T. Shopen (Ed.), Language typology andsyntactic description: Vol. 5. Grammatical categories and thelexicon (pp. 57-144). New York: Cambridge University Press,

Tomasello, M., & Kruger, A. C. (1992). Acquiring verbs inostensive and nonostensive contexts. Journal of Child Lan-guage, 19, 311-333.

Trabasso, T., & Bower, G. H. (1968). Attention in learning: Theoryand research. New York: Wiley.

Tversky, A., & Kahneman, D. (1973). Availability: A heuristic forjudging frequency and probability. Cognitive Psychology, 5,207-232.

Received March 16,1995Revision received June 28,1996

Accepted July 6,1996