kersten 1998 two competing attentional mechanisms in category learning

7/31/2019 Kersten 1998 Two Competing Attentional Mechanisms in Category Learning

1/22

Journal of Experimental Psychology:Learning, Memory, and Cognition1998, Vol. 24, No. 6 ,143 7-14 58Copyright 1998 by the American Psychological Association, Inc.O278-7393/98/S3.00

Two Com peting Attentional M echanisms in Category LearningAlan W. Kersten , Robe rt L. Go ldstone, and Alexan dra SchaffertIn d ian a U n iv er s i ty

This research provides evidence that there are 2 competing attentional mechanisms in categorylearning. Attentional persistence directs attention to attributes previously found to bepredictive, whereas attentional contrast directs attention to attribute values that have notalready been associated with a category. Three experiments provided evidence for thesemechanisms. Experiments 1 and 2 provided evidence for persistence because increasedattention to an attribute followed training in which that attribute was relevant. Theseexperiments also provided evidence for contrast because attention was also increased to thevalues of an attribute when the values of another, more salient attribute had already beenassociated with categories. Experiment 3 provided evidence that persistence operatesprimarily at the level of attributes, whereas contrast operates at the level of attribute values.

Learning theorists have relied on the notion of attention toexplain how learning one discrimination can be facilitated orhindered by the prior learning of another discrimination. Inparticular, learning a discrimination is dependent on atten-tion to the relevant attribute. Two seemingly contradictorymechan isms have been proposed for the changes in attentionfollowing a successful discrimination. According to onemechanism, which we call attentional persistence, increasedattention is allocated to attributes that have been found to bepredictive of an outcome. This mechanism is involved in theanalyzer theory of Sutherland & Mackintosh (1971 ). Accord-ing to the analyzer theory, when a discriminative stimulus inan operant conditioning paradigm is found to be predictiveof reinforcement, greater attention is devoted to the relevan tattribute of the discriminative stimulus. Because the totalpool of attentional resources is fixed, this requires a de creasein attention to other a ttributes. As a result, a learner is betterable to learn a second discrimination involving the sameattribute but is less able to learn a discrimination involving adifferent attribute. Attentional persistence plays a role notonly in theories of operant conditioning but also in promi-nent theories of categorization (e.g., Kruschke, 19 92; Nosof-sky, 1986).

A second attentional mechanism, which we call atten-tional contrast, works in the opposite direction. In particular,once a value of an attribute has been associated with acategory, attention is allocated to other attribute values w henanother category must be learned. This mechanism isinvolved in the Pearce-Hall model of learning (Pearce &Alan W. Kersten, Robert L. G oldstone, and Alexandra Schaifert,Department of Psychology, Indiana University.We thank Julie Earles for helpful comments on a previousversion of this article.Examples of the events used as stimuli in this research areavailable over the World Wide Web at http://www.psy.fau.edu/chez/awk/home.html.Correspondence concerning this article should be addressed toAlan W. Kersten, who is now at the Department of Psychology,Florida Atlantic University, 777 Glades Road, B oca Raton, Florida33431-0991. Electronic mail may be sent to [email protected].

Hall 1980). According to this model, stimuli receiveattention only to the extent that their consequences areunclear. Thus, once an attribute value has been found to bereliably associated with a given outcome, attention to thatattribute value is reduced. This theory thus appears to standin opposition to the analyzer theory, in which a consistentpredictor of an outcome receives greater attention.Eviden ce for Attent ional Pers is tence

Early evidence for attentional persistence can be found inthe classic study of Lawrenc e (19 49). In the first stage of thisstudy, rats in the experimental condition were placed in amaze with one white arm and one black arm. We refer tosuch dimensions of variation, regardless of whether theyvary quantitatively or qualitatively, as attributes (e.g., color).We refer to specific points along these dim ensions as values(e.g., black). Rats were reinforced for entering the blackarm. Attentional persistence would thus predict increasedattention to the color attribute following this d iscrimination.In contrast, rats in a control condition were trained in a graymaze with texture rather than color predicting reinforce-ment. After this first stage of learning, rats were placed inmazes that had either two white arms or two black arms.Rats were reinforced for entering the right arms of blackmazes and the left arms of white mazes. Thus, priorassociations learned in the first stage involving specificvalues of color would not help rats learn these newcontingencies. Consistent with attentional persistence, experi-mental rats learned these contingencies faster than controlrats, presumably because of increased attention to colorfollowing the first stage of training.

Further evidence for persistence has come from studies ofintradimensional versus extradimensional shifts in discrimi-nation. For example, Mackintosh and Little (1969) pre-sented pigeons with pairs of stimuli that varied on both thecolor and orientation of a stripe. One group of pigeons wasreinforced for pecking a stripe of a particular color, whereasa second group was reinforced for pecking a stripe of aparticular orientation. Following this initial training, bothgroups were transferred to a task in which only color was1437


2/22

1438 KERSTEN, GOLDSTONE, AND SCHAFFERTrelevant. Two new colors and two new orientations wereused so that prior associations would presumably notdirectly affect pigeons' performance. Pigeons that weresubject to an intradimensional shift (i.e., those that wereinitially trained on color) performed better on this transfertask than did pigeons who were subject to an extradimen-sional shift. This finding may reflect attentional persistenceto the attribute that was relevant in the first stage of learning.Evidence for persistence in humans has come from astudy by Kruschke (1996). Kruschke first trained undergradu-ate participants on a categorization problem requiring atten-tion to two out of three varying attributes. Some participantswere transferred to a categorization problem requiringattention to one of the previously relevant a ttributes. Otherswere transferred to a problem requiring attention to thepreviously irrelevant attribute. Participants who w ere able tocategorize the transfer stimuli on the basis of a previouslyrelevant attribute learned the categorization more quicklythan those who had to shift attention to a previouslyirrelevant attribute. Recent evidence by Macho (1997) hasalso demonstrated that human participants persist in attend-ing to a previously relevant attribute when a new set ofcategorization stimuli is presented.

Evidence for Attentional ContrastEvidence for contrast comes from word-learning biases inchildren. Children who are presented with a novel categorylabel have been found to attend more to a previouslyunlabeled object than to an already labeled object (M arkman& Wachtel, 1988). Thus, consistent with contrast, childrenattended more to information that was previously unpredic-tive of category membership than to information that they

had already discovered to be predictive.Although this effect is often discussed at the level ofwhole objects, it may be more accurate to characterize it interms of the values of individual attributes. Children havebeen found to associate a novel noun primarily withparticular attributes of an object such as sh ape. In particular,when children are presented with a novel word along with anovel object, they generally extend that word to objects w iththe same shape but not to objects with a different shape(Landau, Smith, & Jones, 1988). For example, a child whohas just learned the word horse would likely extend the useof that label to similarly shaped animals such as zebras butnot to similarly colored animals such as cows. If a novelword is presented along with a shape that already has a label,however, children shift attention to the shapes of otherobjects that are present. For example, if a child werepresented with the word co w and the only objects presentwere a previously labeled horse and an unlabeled cow, thechild would be likely to associate the new word with theshape of the cow. If there are no other objects, children shiftattention to values of other attributes of an object (Markman& Wachtel, 1988). For example, if a child who had justlearned horse was presented with the label zebra and theonly object present was a zebra, that child would be likely toshift attention away from the shape of the object and tow ard

other attributes of the object such as its coloration (i.e., itsstripes).Markman & Wachtel (1988) interpreted these findings asevidence for a mutual exclusivity assumption in children.According to this assumption, an object can have only onelabel, and thus if no unlabeled objects are present, a novellabel must refer to something else. Although mutual exclusiv-ity may reflect a conscious word-learning strategy on thepart of children, it may also reflect the operation of a moregeneral attentional mechanism. Research in the animalconditioning literature has revealed results that are quitesimilar to those in child language learning. For example, inthe phenomenon called "latent inhibition," animals havedifficulty learning associations involving a stimulus towhich they have previously been exposed (Lubow & M oore,1959). For example, after preexposure to a light, rats havedifficulty learning the later predictiveness of the light if boththe light and a novel tone predict reward. According to onemajor account of latent inhibition (Hall, 1991), associationsdeveloped by a stimulus during preexposure interere with itsability to enter into subsequent associations. This samemechanism, which is quite consistent with contrast, couldalso account for a bias to associate novel words with novelreferents in children. In particular, once an attribute valuehas been associated with a category, children may havedifficulty learning new associations involving that attributevalue.

An additional rinding that may reflect contrast has comefrom Schyns and Rodet (1997). Participants learned twocategories of "Ma rtian ce lls," one defined by a single novelfeature (X ) and one defined by this feature contiguous with asecond novel feature (Y). The X-XY group learned the Xcategory prior to the XY category. This group w as predictedto represent the XY category as the conjunction of the X andJY features. In contrast, the XY-X group learned the XYcategory before the X category. This group was expected torepresent the XY category as a single conjoined featurebecause they had never seen X in isolation before learningthe XY category. Participants were tested with stimuliinvolving the same X and Y features separated in space. TheXY-X group tended to categorize these as exam ples of the Xcategory because the conjoined XY feature was not presentbut the X feature was present. In contrast, the X-XY grouptended to categorize them as XY because both the X and Yfeatures w ere present.Schyns and Rodet (1997) interpreted their results asevidence for feature creation. In particular, Schyns and

Rodet proposed that the creation of an X feature in the firstphase of learning allowed participants in the X-XY condi-tion to extract the X feature in the second ph ase of learningand thus to create an independent Y feature. Attentionalcontrast may have worked hand in hand with featurecreation, however, to draw attention to the Y feature in thiscondition. In particular, contrast may have directed attentiontoward information that was not already associated with theX category when a new category was to be learned, resultingin the creation of a Y feature. As a result of this attention toY, the presence of the Y feature in the test stimuli m ay havecaused participants to categorize these stimuli as examples


3/22

COMPETING ATTENTIONAL MECHANISMS 1439of the XY category. In contrast, participants in the XY-Xgroup may ha ve been unable to attend to the absence of Y inorder to distinguish the X category from the previouslylearned XY category. As a result, these participants weremore likely than participants in the other condition tocategorize the test stimuli as X, even though Y was in factpresent in these stimuli.

The In teract ion of Attent ional Pers is tenceand Contras t

Findings consistent with contrast have provided evidencethat learners direct attention away from information that waspreviously found to be predictive of category membership,whereas findings consistent with persistence have providedevidence that learners direct attention toward this previouslypredictive information. The distinction between attributesand attribute values may be important in resolving thisapparent inconsistency. Persistence may operate primarily atthe level of attributes. For example, in the intradimensionalversus extradimensional shift literature, the particular valuesof the attributes that were represented in the transfer stimuliwere all entirely novel. Thus, an attentional mechanism thatoperated only at the level of values could not accoun t for anydifferences in transfer between the intradimensional andextradimensional shift conditions. The intradimensional con-dition, however, had received prior training on the relevantattribute, and this resulted in better transfer performance.Thus, when an attribute was found to be predictive of anoutcome, persistence resulted in increased attention to thatattribute regardless of the values involved.

Contrast may operate primarily at the level of attributevalues. Studies that have provided evidence for contrasthave all used transfer stimuli involving values of an attributethat have previously been associated with a category. Forexam ple, children's association of a word (e.g., horse) with aparticular attribute value (e.g., a horse shape) encouragesthem to focus on different values (e.g., a cow shape) whenpresented with a novel label (e.g., cow). In this example,contrast does not direct attention away from an entireattribute (e.g., shape) but rather from a particular value ofthat attribute (e.g., a horse shape). Contrast does not requireattention to be shifted to another value of the sam e attribute,however. For example, if no novel shapes are present,contrast results in increased attention to values of otherattributes such as the substance (Markman & Wachtel, 1988)or motion (Kersten, Fancher, Remec, & Schaefer, 1997) ofan object. Thus, although contrast encourages attention tovalues that have not already been associated with a category,it is indifferent as to which attributes those values comefrom.

Persistence may direct attention to an attribute, whereascontrast directs attention away from p articular values of thatattribute that already have a label. For exam ple, a child whois presented with a novel category label along with apreviously labeled shape will be likely to continue to attendto shape if other, unlabeled shapes are present. Smith (19 95)has presented evidence that this shape bias is not innate butrather develops with noun-learning experience, consistent

with an attentional mechanism such as persistence. Thus,after learning the word horse, persistence would directattention to shape, whereas contrast would direct attentionaway from horselike shapes. As a result, a child who wasthen presented with the novel word co w along with a cowand a zebra would be likely to associate the word with theshape of the cow rather than with the coloration of the zebra.Only if no unlabeled values of an attribute are present doescontrast require attention to be shifted to the values of adifferent attribute. In situations such as this, attentionalpersistence and contrast may compete with one another todetermine the allocation of attention, with persistence favor-ing one attribute (e.g., shape) and contrast directing attentionto the values of a different attribute (e.g., coloration).When attentional persistence and contrast are in competi-tion with one another, a key factor in determining whichprocess exerts a stronger influence may be the baseline levelof attention paid to each attribute. In a recent extension of bisattention learning covering map (ALCOVE) model, Krus-chke (1996) has proposed that the prior level of attention to

an attribute may influence the degree of chang es in attentionto that attribute. In particular, attributes that are currentlyreceiving very high levels of attention show smaller changesin attention strength than do attributes that are currentlyreceiving more moderate levels of attention. Thus, anattribute that normally receives a high level of attention mayreceive little extra attention as a result of either persistenceor contrast. For example, after children have learned wordsfor a number of different shapes, they may attend strongly toshape as a result of persistence. Because of this already highlevel of attention, however, little extra attention will accrueto shape following the learning of each additional word for ashape.An attribute that normally receives a more moderate levelof attention may receive a greater increase in attention as aresult of persistence or contrast. Thus, if one attentionalprocess favored a highly attended attribute and the otherfavored a more moderately attended attribute, the processfavoring the m ore moderately attended attribute would exerta stronger influence. For example, if persistence favored ahighly attended attribute such as shape and contrast favoreda more moderately attended attribute such as coloration,contrast would e xert a larger influence, resulting in a relativeincrease in attention to coloration and a relative decrease inattention to shape.The present experiments were designed to examine theinfluences of persistence and contrast on attention to at-

tributes that varied in baseline salience. Participants learneda discrimination in which the values of either a more or lesssalient attribute could be used to distinguish betweencategories. Some participants learned this discrimination atthe beginning of the experiment, prior to learning any otherdiscriminations. These participants were expected to attendprimarily to the more salient attribute when making thisdiscrimination. Other participants first learned a discrimina -tion in which the same values of only one of these twoattributes could be used to distinguish between categories.Thus, persistence would result in increased attention to thisattribute in the learning of subsequent discriminations.


4/22

1440 KERSTEN, GOLDSTONE, AND SCHAFFERTContrast would result in increased attention to the otherattribute because the values of this attribute would not yethave been associated with categories.Because persistence and contrast worked in oppositedirections in this design, effects of prior learning would beapparent only if one attentional mechanism had strongereffects than the other. As we discussed previously, thebaseline salience of an attribute may be important indetermining the degree of change in attention to an attribute.Because persistence and contrast favored attributes thatdiffered in baseline salience, the mechanism that favored theless salient attribute was expected to yield a strongerinfluence. Persistence was expected to dominate over con-trast when the less salient attribute was involved in the initialdiscrimination, resulting in increased attention to that at-tribute in subsequent discriminations. Contrast was expectedto dominate over persistence when the more salient attributewas involved in the initial discrimination, again resulting inincreased attention to the less salient attribute in subsequentdiscriminations.

Exper iment 1Experime nt 1 was designed to exam ine the effects ofattentional persistence and contrast on attention to attributesthat varied in baseline salience. The attributes that definedthe categories were components of simple animated events.Novel verbs accompanied the events, serving as labels forthe event categories. Event categories were chosen forinvestigating attentional persistence and contrast becausedifferent attributes of events show consistent differences intheir baseline saliences when they are presented in thecontext of a verb-learning task (Kersten, 1997, 1998, inpress). They were also chosen because the present contrastmechanism may be particularly important in the learning ofEnglish verbs (see the General Discussion). We believe thatthe proposed attentional mechanisms are quite general innature, however, and would also apply to the learning ofobject categories.Two attributes of the events were related to the accom pa-nying verbs. These were the path and manner of motion ofone of the characters appearing in each event. Theseattributes were chosen because peop le have previously beenfound to attend more to path than to m anner of motion in thecontext of a verb-learning task (Kersten, 1997). In otherwords, path has a higher baseline salience than manner ofmotion. Each verb was always accompanied by a particular

value of one or both of these attributes. For example, oneverb was always accompanied by a character moving alonga path directly toward a second character, whereas anotherverb was always accompanied by a manner of motioninvolving zig-zagging to the left and right.The order of learning different types of verbs was variedto influence the amount of attention allocated to eachattribute. To provide a baseline m easure of attention to pathand manner of motion, one group of participants learnedverbs that could be distinguished on the basis of either pathor manner of motion before learning any other discrimina-tions. For example, one such verb was always accompanied

by both the path and manner of motion described above,whereas a second verb was always accompanied by adifferent path and manner of motion. Because these verbswere defined by a conjunction of values of path and mannerof motion, they are referred to as conjunctive verbs. Becausepath is more salient than manner of motion, participants wholearned conjunctive verbs first were expected to attendprimarily to path when learning these verbs.Other participants learned verbs that could only bedistinguished on the basis of manner of motion beforelearning conjunctive verbs that could be distinguished on thebasis of path as well as these same values of manner ofmotion. These participants we re predicted to attend more tomanner of motion when learning conjunctive verbs thanparticipants who learned conjunctive verbs first. This findingwould provide evidence for attentional persistence. Becausemanner of motion was initially less salient than path, thispersistence to manner of motion was expected to overwhelmany contrast effects acting to draw attention to path.The remaining participants learned verbs that could only

be distinguished on the basis of path before learningconjunctive verbs that could be distinguished on the basis ofmanner of motion as well as these same values of path.These participants were also predicted to attend more tomanner of motion when learning conjunctive verbs thanparticipants who learned conjunctive verbs first. This findingwould provide evidence for contrast. This effect of contraston attention to manner of motion was expected to over-whelm any persistence effects acting to draw attention topath. Thus, prior learning of verbs involving either path ormanner of motion was expected to result in an increase inattention to manner of motion during subsequent conjunc-tive learning.Although the order of learning varied, all participantslearned path verbs, manner verbs, and conjunctive verbs.For example, participants who learned conjunctive verbsfirst subsequently learned path verbs and manner verbs.Thus, at the end of learning, all participants could be testedon their ability to distinguish between these different verbtypes. A participant's ability to distinguish conjunctive verbsfrom path verbs provided a measure of attention to mannerduring conjunctive learning because only manner distin-guished these two verb types. A participant's ability todistinguish conjunctive verbs from manner verbs, in turn,provided a measure of attention to path.Participants who learned conjunctive verbs first wereexpected to attend primarily to path. As a result, these

participants were expected to have difficulty distinguishingconjunctive verbs from path verbs. Participants who learnedpath or manner verbs first were expected to attend more tomanner during subsequent conjunctive learning than partici-pants who learned conjunctive verbs first. Because manneris lower in baseline salience than path, this increase inattention to manner following path or manner learning wasexpected to result in a more even distribution of attentionover path and manner. These participants were thus morelikely to notice the relation of conjunctive verbs to both pathand manner, allowing them to distinguish conjunctive verbsfrom path verbs.


5/22

COMPETING ATTENTIONAL MECHANISMS 1441Method

ParticipantsOne hundred forty-four undergraduates at Indiana Universityparticipate d in this experim ent in partial fulfillment of courserequirements for introductory psychology.

morping

StimuliAll events. Events were displayed on Macintosh Dsi (AppleCom puters, Cupertino, CA) computers using MacroMind Director3.1 (Macromedia, San Francisco, CA). An example event isdepicted in Figure I. E ach event involved two buglike characters.One character, the agent, moved throughout the event, while theother, the patient, remained motionless. The static appearance ofeach character varied on three attributes, its head, body, and legs.Each of these attributes had tw o possible values, chosen randomlyfrom the four values used by Kersten (1998). An agent's motionalso varied on three attributes. These were the path, manner ofmotion, an d leg motion of the agent. The two possible values ofeach of these attributes are depicted in Figure 2. The only other

attribute that varied was the environment in which the events took,place, represented by a static line-drawing in an unused corner ofthe screen. There were two possible environments, chosen ran-domly from the four environments used previously by Kersten(1998).Each event began after a black screen faded away to reveal thestarting positions of the two characters. The agent started eachevent at a point randomly chosen from a region near the center ofthe screen. The patient started each event at a varying distance fromthe agent along one of eight directions, either north, south, east,west, northeast, northwest, southeast, or southwest. Each eventlasted approximately 8 s.Learning events. There were three phases of learning. Partici-pants learned path verbs in one phase, manner verbs in another, andconjunctive verbs in a third. Two verbs were presented in each

phase of learning. Each phase involved between one and threeblocks of eight learning events. Each verb was presented on four ofthe eight events within a block, with the order of presentationdetermined randomly for each participant. Participants were testedon their learning following each block. They were allowed to m oveon to the next phase of learning if testing indicated that they hadlearned the verbs in that phase. Participants were allowed to moveon regardless of test performance following three blocks oflearning in a given phase.Table 1 shows a schematic description of the categories inExperiment 1. Each event w as labeled by a novel verb appearing inan unused corner of the screen. In the path phase, the two verbs(i.e., spogging an d zelling) were perfectly correlated with the twodifferent paths of the agent. Manner of motion did not vary in thisphase, with agents moving smoothly throughout. In the mannerphase, the two verbs (i.e., yimming and gupping) were perfectlycorrelated with the two different manners of motion of the agent.Path did not vary in this phase, with agents oscillating in placeeither front and back or left and right to represent their manners ofmotion. In the conjunctive phase, the two verbs (i.e., morping andwunking) were perfectly correlated both with the paths and with themanners of motion of the agent. All other attributes variedrandomly. This random variation was included to make the taskmore challenging.Criterion test trials. Following each block of eight learningevents, participants were tested on their knowledge of the twoverbs they just saw. This was done to ensure that participants hadlearned the meanings of those verbs before allowing them to move

morping

morping

Figure 1. Three frames from an example event in Experiment 1.The agent, on the right, zigzags to its left and right as it movestoward the patient, on the left. The agent's legs move in a rotatingmotion around its body as it move s. The event is labeled by the verbmorping and takes place against a desert background.

on to the next phase of the experiment. The amount of difficultyparticipants experienced when learning a particular discriminationcan also be used as a measure of attention to the relevantattribute(s) in that discrimination. There were six test trialsfollowing each block of learning events, three testing for knowl-edge of each verb. The order of presentation of the test trials wasdetermined randomly for each participant. Participants were pre-sented with pairs of events, one after the other, each labeled by thesame verb. One event in each pair was consistent with theaccompanying verb, whereas the other mismatched the verb oneither path (in the path phase), manner of motion (in the mannerphase), or both (in the conjunctive phase). The correct event waspresented first in half of the trials and second in the other half. Th e


6/22

1442 KERSTEN, GOLDSTONE, AND SCHAFFERTP a t h

Manner of Motion

w

Leg Motion

Figure 2. Values of the motion attributes in Experiment 1. Onepath of the agent involved motion directly toward the patient,whereas the second involved an indirect path to the patient. Onemanner of motion involved accelerations and decelerations inforward motion (i.e., moving in surges), whereas the secondinvolved leftward and rightward excursions from the agent's path(i.e., zig-zagging). The two leg motions involved rotating motions45 forward and backward around the body of the agent, either inthe same direction by the two legs or in opposite directions.

participant's task was to choose which event was the betterexample of the accompanying verb. The criterion for moving on tothe next phase of the experiment was 100% correct performance onthe set of six test trials. Participants were allowed to move onregardless of performance after three blocks in a given phase oflearning.Final test events. Following the third phase of learning,participants Were presented with 20 events testing for know ledge ofall six verbs seen earlier in the experiment. These final test eventswere used to determine how well participants could distinguishverbs that were learned in different phases of learning. Recall that

conjunctive verbs differed from each of the other two verb typ es ononly one attribute. Thus, a participant's ability to distinguishconjunctive verbs from another verb type (e.g., path verbs)provided an estimate of attention during conjunctive learning to theattribute that distinguished those two verb types (e.g., manner).Table 1 presents schem atic desc riptions of the five different testtypes. There were four trials of each of the five test types. The firstthree typespath, manner of motion, and conjunctive trialsinvolved events that were generated in the same way as events seenduring learning. After each event, participants were asked tochoose w hich of the six verbs was the best label for that event. Oneof the two verbs from the corresponding phase of learning was thusthe correct answer in these trials. The fourth type, mismatch trials,involved events that depicted a combination of path and manner ofmotion that had not been presented during learning.The fifth test type, the separated trials, was designed to ensurethat participants represented path and manner as independentattributes during conjunctive learning. Because path and mannerwere perfectly correlated during conjunctive learning, participantscould have represented path and manner as a single, conjoinedattribute. For example, participants could have construed the pathand the manner in the left colum n of Figure 2 as a zig-zagging path.If each manner of motion were subsumed into a complex path inthis way, then participants could have defined conjunctive verbs interms of these complex paths and differentiated conjunctive verbsfrom path verbs without reference to an independent manner ofmotion attribute.

The separated trials used a method analogous to that of Schynsand Rodet (1997) to test whether participants could distinguishconjunctive verbs from path verbs on the basis of an independentmanner-of-motion attribute. In particular, path and manner ofmotion were presented as separate entities. Whereas the features inSchyns and Rodet's research were separated spatially, path andmanner were separated in the present work by using two differentobjects to carry these two different types of motion. In particular,the agent carried the manner-of-motion attribute, just as it hadthroughout learning. The agent began the event oscillating in place,just as during the manner phase of learning. While it did this,however, a flying carpet, represented by a red square, flewunderneath the agent and then carried the agent along on its path.The carpet flew along one of the two paths taken by the agentduring learning. The agent started the event about halfway alongthe path from the carpet to the patient, thus allowing the carpet tostart in motion independently of the agent and then pick it up. This

Table 1Schematic Description of Stimuli From the Three Phases ofLearning and From the Final Test Trials of Experiment 1Event type

Learning phasePathMannerConjunctiveFinal testPathMannerConjunctiveMismatchSeparated

Stimulus 1P . MoP o M ,P , M ,P | M oP o M ,Pi M,P , M 2P, (Carpet) Mi

Stimulus 2P 2M oP 0 M 2P 2 M 2P2M0P 0 M 2P 2 M 2P 2 M ,P 2 (Carpet) M 2

Note. Subscript numbers refer to the values of these attributes. Asubscript 0 next to an attribute indicates that there was no variationon that attribute . The term Carpet after a path indicates that the pathwas taken by a flying carpet. P = path; M = manner.


7/22

COMPETING ATTENTIONAL MECHANISMS 1443was done to encourage the interpretation that the agent was notresponsible for its own motion but rather was being carried by thecarpet. The agent, however, continued to oscillate throughout theevent while the carpet moved smoothly. The carpet thus repre-sented the path attribute, whereas the agent represented manner ofmotion. The path and manner of motion were paired as they hadbeen during the conjunctive phase of learning. Thus, a participantwho represented conjunctive verbs in terms of separate path andmanner attributes would be expected to choose conjunctive verbsas labels for the separated trials.Procedure

Participants were instructed that they would be learning verbsfrom a language spoken on another planet. They were thenpresented w ith the first phase of learning. After each even t, a buttonlabeled Next Event appeared in the lower right-hand corner of th escreen, allowing participants to continue on to the next e vent. Aftereight learning events, participants were given sixcriterion testtrials. After the first event in each test trial, participants clicked onth e Next Event button to see the second event. After the secondevent, participants clicked on one of two buttons labeled FirstEvent an d Second Event to indicate which event they thought was abetter example of the verb accompanying those events. A thirdbutton labeled Repeat allowed participants to review the tw o eventsas many times as they wished. If participants were correct on all sixevents, they were allowed to continue on to the next phase oflearning. If participants failed to achieve this criterion, theyrepeated the learning and test cycle up to two more times. Thissame procedure was followed for all three phases of learning.Different participants were presented with the three phases oflearning in sixdifferent orders, representing all possible permuta-tions of the path, manner, and conjunctive phases of learning. Aftercompleting the three phases of learning, participants were shown20 final test trials. At the end of each event, participants chose oneof six buttons, each labeled by an individual verb, to indicate whichverb they thought was the best label for that event. The Repeatbutton also appeared, allowing participants to review an event asoften as they wished.

DesignThe primary dependent variable in this experiment was thepercentage of correct conjunctive responses in the conjunctive andseparated trials. The independent variables for this analysis werethe order in which participants received the three phases oflearning, manipulated between participants, and the type of test

(conjunctive vs. separated), manipulated within participants. Addi-tional dependent variables were the percentages of path and m annerresponses to the path, manner, andmismatch trials. The indepen-dent variable for these measures was learning order. A finaldependent variable was the percent correct on criterion test events.The independent variables for this dependent variable were learn-ing order and the type of verb being tested (i.e., path vs. manner v s.conjunctive).

ResultsCriterion Test Trials

The results of the first block of criterion test events witheach verb type are displayed in Figure 3. The relativedifficulty of the different verb types provides a measure ofattention to path and manner of motion. Because participantswho performed with perfect accuracy in a block of criteriontrials were moved on to the next phase of the experiment, notall participants completed three blocks of learning for eachverb type (i.e., path, manner, and conjunctive). All partici-pants completed at least one block of trials for each verbtype, however, and thus an analysis of variance (ANOVA)could be performed on criterion test scores following thefirst block of learning with each verb type.

The effect of verb type was analyzed using two plannedcomparisons. The first comparison tested whether partici-pants performed better when they could use path to make adiscrimination, compared with when they could only usemanner. In particular, performance with path verbs and

100S 90CD!_i-oocQ )oI0)

8070 -6050 Conjunctive Path

Verb

IM a n n e r

P -> M -> CM -> P -> CP -> C -> MM -> C -> PC -> P -> MC -> M -> P

T y peFigure 3. Results of the first bloc k of criterion test trials foreach verb type in Experiment 1. The sixconditions reflect the six orders of learning (e.g., P > M C = path manner conjunctive).Error bars reflect standard errors.


8/22

1444 KERSTEN, GOLDSTONE, AND SCHAFFERTconjunctive verbs was compared with performance withmanner ve rbs. This comparison revealed worse performancewith manner verbs than with the other two verb types, F(l,138) = 2 3.65,/? < .001, M SE = 628.37 , suggesting that pathis indeed more salient than manner. The second co mparisontested whether participants performed better when theycould use both path and manner to make a discrimination,compared with when they could only use path. In particular,performance with conjunctive verbs was compared withperformance with path verbs. This comparison revealed nodifferences between the two verb types, F(l, 138) < 1,suggesting that relevant variation on manner added littlediscriminability to conjunctive verbs above and beyond thatprovided by path.

There was also a significant effect of learning order oncriterion test scores, F(5, 138) = 4.95, p < .001, M SE =345.74, as well as a significant interaction of learning orderwith the comparison between manner verbs and the othertwo verb types, F(5, 138 ) = 4.97,/? < .001, MSE = 628.37.The comparison between path verbs and conjunctive verbs,on the other hand, was not dependent on learning order, F(5 ,138) < 1. To unde rstand the cau se of the significant effectsinvolving learning order, post hoc Fisher's least significantdifference (LSD) analyses compared the six learning orderson each verb type in isolation. These analyses revealed nosignificant differences on path and conjunctive verbs. Therewere, however, significant differences on manner verbs. Inparticular, participants in the Path-Manner-Conjunctive(PMC) and PCM conditions learned manner verbs morequickly than participants in the MPC, MCP, and CMPconditions; participants in the CPM condition performedbetter than participants in the MPC and MCP cond itions; andparticipants in the CMP condition performed better thanparticipants in the MPC con dition. Thus, participants appearto have been better at learning manner verbs if they hadalready learned path verbs, compared with when theylearned m anner verbs first. This finding may reflect a pull onattention from novel variation on manner. In particular,manner did not vary when participants learned path verbs.When participants subsequently learned manner verbs, novelvariation on manner was introduced. This novel variationmay have drawn attention to manner, allowing participantsto associate verbs with manners of m otion.

Final Test TrialsParticipants were selected for analysis of final test scoreson the basis of performance in the criterion test trials. Inparticular, participants had to perform with 100% accuracyon at least one block of criterion test trials for each verb typeto be included in analyses of final test scores. Out of the 2 4participants in each condition, the num bers of participants ineach condition who achieved this criterion were as follows:24 in the PMC condition, 17 in the MPC cond ition, 24 in thePCM condition, 18 in the MCP condition, 24 in the CPMcondition, and 18 in the CMP condition.The percentages of correct conjunctive choices in theconjunctive and separated trials are depicted in Figure 4.These trials tested participants' abilities to discriminate

conjunctive verbs from path verbs, providing a measure ofattention to manner of motion during conjunctive learning.An ANOVA with trial type and learning order as indepen-dent variables revealed a m ain effect of learning order, F(5 ,119) = 4.98, p < .001, M SE = 1,163.01. As predicted,participants who learned path or manner verbs prior toconjunctive verbs performed better than participants wholearned conjunctive v erbs first.

There was also a main effect of trial type, F{\, 119) =62.79, p < .001, M SE = 822.45, with more correctconjunctive choices in the conjunctive trials than in theseparated trials. This may indicate that some participants didnot represent conjunctive verbs in terms of independent pathand manner attributes. If they instead represented path andmanner as a conjoined attribute, they may have rejectedconjunctive verbs as labels for the separated trials becausepath and manner were portrayed separately in these trials.The interaction of Trial Type X Learning Order, however,did not approach significance, F(5, 119) = 0.33, p > .10,MSE = 822.45. Thus, any effects of learning order in thisexperiment canno t be explained by differences in representa-tion because these effects were similar in the separated trialsand in the conjunctive trials. These effects of learning ordermust therefore be a result of attentional me chanisms.

To test for effects of the two propo sed attentional m echa-nisms on performance in the conjunctive and separatedtrials, we created a composite score of conjunctive verblearning by averaging each participant's z scores for thesetwo measures. Two nonorthogonal planned comparisonswere then carried out on these composite scores to test foreffects of persistence and contrast on the learning ofconjunctive verbs. An alpha level of .025 was adopted foreach of these comparisons to m aintain an overall alpha levelof .05.One comparison compared the PMC and MCP conditionswith the PCM and CMP conditions, testing for effects ofpersistence on attention to manner during conjunctive learn-ing. Each of the former conditions is identical to one of thelatter conditions except that the order of learning mannerand conjunctive verbs is reversed. Participants who learnedmanner verbs first were expected to subsequently attendmore to manner in conjunctive learning as a result ofpersistence. The CPM and MPC conditions were not in-cluded in this comparison because path learning intervenedbetween manner and conjunctive learning. As a result, itwould have been impossible to determine whether changesin attention to manner were due to persistence, contrast, orboth. This comparison was significant, r(119) = 2.43, p


9/22

COMPETING ATTENTIONAL MECHANISMS 1445

Conjunctive Trials

100

ooo0)o%a >



S e p a r a t e d T r i a l sFigure 4. Percentage of coirect conjunctive choices in the conjunctive and separatedfinal est trialsof Experiment 1. The six conditions reflect the six orders of learning (e.g., P M C = path manner * conjunctive). Error bars reflect standard errors.

path verbs first were expected to attend more to manner insubsequent conjunctive verb learning as a result of contrast.The CMP and PMC conditions were not included in thiscomparison because manner learning intervened betweenpath and conjunctive learning. This comparison was alsosignificant, r(119) = 2.34, p < .025, providing evidence forcontrast. In particular, participants who learned path verbsprior to conjunctive verbs made more correct conjunctivechoices in the conjunctive and separated trials than didparticipants who learned conjunctive verbs before pathverbs.An analysis of incorrect responses in the conjunctive andseparated trials revealed that participants who failed tochoose correct conjunctive verbs most often chose a pathverb. Such verbs conveyed the same paths as conjunctiveverbs but did not convey the same m anners of motion as did

conjunctive verbs. Thus, a participant who c hose a path verbinstead of a conjunctive verb in these trials failed to noticethe manner component of conjunctive verb meanings. Thepercentages of choices in the conjunctive and separatedtrials that matched the displayed event on path but mis-matched on manner of motion are depicted in Figure S. AnANOVA on these percentages revealed a significant maineffect of test type, F ( l , 119) = 5.98,/? < .05, M SE =481.44,with more path choices in the conjunctive trials than in theseparated trials. There was also a significant main effect oflearning order, F(5 ,11 9) = 4.93, p < .001, MSE = 805.25.Participants who learned path or manner verbs beforelearning conjunctive verbs made fewer path verb choicesthan didparticipants who learned conjunctive verbs first.Test type again did not interact with learning order, F(5,119) = 0.70, p > .10, M SE = 481.44.


10/22

1446 KERSTEN, GOLDSTONE, AND SCHAFFERT100

Conjunctive Trials

10003OO

O

TOQ.Co

8060402 0 -o -

m

S e p a r a t e d T r i a l s



Figure 5. Percentage of choices in the conjunctive and separated final test trials of Experiment 1involving path verbs that matched the test event on path but failed to predict the manner of motiondisplayed in the event The six conditions reflect the six orders of learning (e.g., P * M C =path manner conjunctive). Error bars reflect standard errors.

To understand the main effect of learning order on th epercentages of path choices in the conjunctive and separatedtrials, we created a composite score of path verb choices byaveraging each participant's z scores for path verb choices inthe two trial types. The same two planned comparisons thatate described earlier were performed to test for effects ofpersistence and contrast on these composite scores. Thecomparison testing for evidence of persistence examinedwhether participants who learned manner verbs prior toconjunctive verbs made fewer path verb choices in theconjunctive and separated trials. This prior learning ofmanner verbs was expected to increase attention to mannerin conjunctive learning, allowing participants to betterdiscriminate conjunctive verbs from path verbs. This effectof persistence approached significance, r(119) = 1.71, p M -> CM -> P -> CP -> C -> MM -> C -> PC -> P -> MC -> M -> P

P a th T r i a l sFigure 6. Percentage of correct path choices in the path final test trials of Experiment 1. The sixcondition s reflect the six orders of lea rning (e.g., P M C = p ath manner conjunctive).Error bars reflect standard errors.


12/22

1448 KERSTEN, GOLDSTONE, AND SCHAFFERTfrom the criterion trials, in which participants were betterable to learn manner verbs after they had already learnedpath verbs. Novel variation on manner was introduced whenparticipants were transferred from path learning to mannerlearning, perhaps drawing attention to manner and thusallowing participants to better learn manner verbs. Toprovide evidence for contrast independently of any effects ofnovel variation on an attribute, Experiment 2 was designedto replicate the results of Experiment 1 while maintainingthe amount of variation displayed by an attribute throughoutlearning.

Experiment 2Experiment 2 differed from Experiment 1 in that all

attributes that were not relevant to the meaning of a verbvaried randomly during learning. For example, in the pathphase of learning, manner varied independently of theaccompanying verbs. Thus, no new variation on manner wasadded when participants were transferred from the pathphase to the conjunctive phase of learning.A second aim of Experiment 2 was to replicate the resultsof Experiment 1 but using a different type of motion tooperationalize the less salient attribute. In particular, mannerof motion acted as the less salient attribute for half of theparticipants and leg motion played this role for the otherhalf. Kersten (1997) has proposed that manner of motion andleg motion are similar in that they are both examples ofintrinsic motion, that is, motion defined with respect to aframe of reference internal to the object carrying out thatmotion. Prior experiments have revealed that the present legmotions, similar to the present manners of motion, receiveless attention than the present paths in the context of verblearning (Kersten, 1998; Kersten & Billman, 1995). Ifpersistence and contrast are general attentional processes,they should be able to draw attention to an attribute of lowbaseline salience regardless of the identity of that attribute.We thus predicted similar effects for leg motion and mannerof motion.

Table 2Schematic Description of Stimuli From the Three Phasesof Learning in Experiment 2

MethodParticipants

One hundred sixty-two undergraduates at Indiana Universitytook part in this experiment in partial fulfillment of courserequirements for introductory psychology.

StimuliLearning events. Table 2 shows a schematic description of th ecategories in Experiment 2. The learning events of Experiment 2differed from those of Experiment 1 in that for half of theparticipants, legmotion rather than manner of motion was relatedto verb meaning. Manner of motion was relevant for the other halfof the participants, as in Experiment 1. Because these two differenttypes of intrinsic m otion were used interchangeably, thephase oflearning in which they were relevant is referred to as the intrinsicmotion phase. In contrast to Experiment 1, in which manner ofmotion and path did not vary when only the other was relevant, in

Learning phase Verb 1 stimulus Verb 2 stimulus

Pi L 12 M 12Pl2 Ll2 M iP i L , 2 M ,P2 L12 M 12P 1 2 L 1 2 M 2P 2 L 1 2 M 2

Intrinsic motion operationalized by leg motionPath Pi L t 2 M12 P2 L12 M12Intrinsic motion P 12 Lj M 12 P i 2 L2 M 12Conjunctive P( L\ M i 2 P 2 L 2 M ( 2

Intrinsic motion operationalized by manner of motionPathintrinsic motionConjunctive

Note. Subscript numbers refer to the values of these attributes. Asubscript 12 following an attribute indicates that th e attribute wasnondiagno stic, taking each value in half of the examples of a verb.P = path; L = le g motion; M = manner of motion.

Experiment 2, path, manner of motion, and leg motion variedrandomly when they were not relevant to verb meaning.Criterion test trials. These differed from tho se of Experiment 1in that there were only four criterion test trials after each block oflearning events. All attributes that were not related to verb meaningin a particular phase of learning varied randomly across the twoevents in a test trial.Final test events. The final test events were identical to thoseof Experiment 1 except that associations involving leg motionrather than manner of motion were tested for half of the partici-pants. Trials testing for knowledge of either leg motion or mannerof motion are referred to as the intrinsic motion trials.Although path had varied randomly during the intrinsic motionphase of learning, path was not displayed in the intrinsic motiontrials during the final test. This was done to ensure that only anintrinsic motion ve rb could be used as an appropriate label for thesetrials. Because all attributes that were not relevant to the meaningof a particular verb varied randomly during learning, the samecombination of values of path, legmotion, and manner of motioncould have been labeled not only by an intrinsic motion verb in theintrinsic motion ph ase of learning but also by a path verb in the pathphase and a conjunctive verb in the conjunctive phase. As a result,if a random path had been chosen for each intrinsic motion trial,then participants could have correctly labeled each such event witheither a path verb, an intrinsic motion verb, or a conjunctive verb.Because path was instead not displayed in the intrinsic motiontrials, a participant who associated a verb with path should not haveused this verb as a label for these trials. Thus, the percentage ofconjunctive choices in the intrinsic motion trials could be used as ameasure of attention to path during conjunctive learning.

Similarly, during the path trials, manner of motion was notdisplayed to participants for whom manner of motion was relevantduring learning, whereas leg motion was not displayed to partici-pants for whom leg motion was relevant during learning. As aresult, a participant who associated a verb with intrinsic motionshould nothave used this verb as a label for the path trials. Thus,the percentage of conjunctive choices in the path trials could beused as a measure of attention to intrinsic motion during conjunc-tive learning.As in Experiment 1, the conjunctive, separated, and mismatchtrials displayed particular combinations of values of path and therelevant intrinsic motion attribute. Because each such combinationcould have been labeled during learning by a path verb and anintrinsic motion verb, as well as by a conjunctive verb, there was no


13/22

COMPETING ATTENTIONAL MECHANISMS 1449longer a correct answer in these trials. As a result, they were nolonger diagnostic of attention to path andmanner during conjunc-tive learning. They were included simply to retain as much aspossible of the test procedure of Experiment 1.Procedure

The procedure was identical to that of Experiment 1 except thatthere were only three learning orders: path-intrinsic motion-conjunctive (PIC), path-conjunctive-intrinsic motion (PCI), andconjunctive-path-intrinsic motion (CPI). In addition, only fourcriterion test trials followed each block of learning.Design

The primary dependent variable in this experiment was thepercentage of conjunctive responses in the path trials. This variablecould be used as a measure of attention to intrinsic motion becauseparticipants w ould be able to reject conjunctive verbs andchoosepath verbs in these trials to the extent that they noticed the intrinsicmotion components of conjunctive verbs. The independent vari-ables for this analysis were the order in which participants receivedthe three phases of learning (PIC vs. PCI vs. CPI) and the attributeused to operationalize intrinsic motion (leg motion vs. manner ofmotion). Both were manipulated between participants. A furtherdependent variable was the percent correct on criterion test events.The independent variables for this dependent variable were learn-ing order and the attribute used to operationalize intrinsic motion,both manipulated between participants, as well as the type of verbbeing tested (i.e., path vs. intrinsic motion vs. conjunctive), whichwas manipulated w ithin participants.


The results of the first block of criterion test trials witheach verb type are depicted in Figure 7. As inExperiment 1,the relative difficulty of the different verb types provides a

measure of attention to path and intrinsic motion. Becauseall participants completed at least one block of learning foreach verb type, the results of trials following the first blockof learning with each verb type were analyzed in anANOVA. The effect of verb type was again analyzed byusing two planned comparisons. The first comparison re-vealed worse performance with intrinsic motion verbs thanwith path verbs and conjunctive verbs, F( l, 156) = 127.77,p < .001, MSE = 858.36. The second comparison revealedno significant differences between path verbs and conjunc-tive verbs, F(l, 156) < 1. There were no main effects orinteractions involving learning order or the attribute used tooperationalize intrinsic motion (all Fs < 1.70). These resultswere consistent with the work of Kersten (1998) and Kerstenand Billman (1995) in showing that both leg motion andmanner of motion are less salient than path in the context ofverb learning.Final Test Events

As in Experiment 1, participants were selected for analy-sis of final test scores if they performed with 100% accuracyon at least one b lock of criterion test trials for each verb type.Ou t of 27 participants in each condition, the numbers ofparticipants in each condition who achieved this criterionwere as follows: 14 in the PIC condition with manner ofmotion, 19 in the PIC condition with leg motion, 13 in thePCI condition with manner of motion, 19 in the PCIcondition with leg motion, 13 in the CPI condition withmanner of motion, and 17 in the CPI condition with legmotion.The percentages of conjunctive choices in the path trialsare depicted in Figure 8. These conjunctive verbs wereconsistent with the presented paths, but they had alwaysbeen accompanied by particular intrinsic motions duringlearning that were not present in the path trials. Thus, the

100

P -> M -> CP -> L -> CP -> C -> MP -> C -> LC -> P -> MC -> P -> L

Conjunctive Path IntrinsicVerb Type

MotionFigure 7. Results of the first block of criterion test trials for each verb type in Experiment 2. Errorbars reflect standard errors. P = path; M = manner; C = conjunctive; L = leg motion.


14/22

1450 KERSTEN, GOLDSTONE, AND SCHAFFERT

P -> I -> CP -> C -> IC -> P -> I

Ma nner Leg Mot ionIn t r i ns i c Mot i on A t t r i bu te

Figure 8. The percentages of conjunctive choices in the pathfinal est trials of Experiment 2 thatmatched the presented event on its path component but mismatched on intrinsic motion. Error barsreflect standard errors. PIC = path > intrinsic motion (i.e., manner of motion or leg motion) conjunctive; PCI = path conjunctive * intrinsic motion; CPI = conjunctive path > intrinsicmotion.

extent to which participants chose conjunctive verbs ratherthan path verbs as labels for these trials can be used as ameasure of attention to intrinsic motion during conjunctivelearning. An ANOVA with learning order and intrinsicmotion attribute as independent variables revealed a maineffect of learning order, F(2,89) = 1 1 . 7 8 , / J < .001, M SE =816.57. There was no main effect of the choice of intrinsicmotion attribute, F(l, 89) = 2.32, p > .10, MSE = 816.57,nor did this variable interact with learning order, F ( l , 89 ) =1.74, p> . 10, MSE= 816.57.Two nonorthogonal planned comparisons were carriedout to understand the main effect of learning order on thepercentages of conjunctive choices in the path trials. Thesecomparisons tested for effects of persistence and contrast onattention to intrinsic motion during conjunctive learning. Analpha level of .025 was adopted for each of these analyses topreserve an overall alpha level of .05.One comparison tested for contrast effects on attention tointrinsic motion during conjunctive learning as a result ofprior path learning. In particular, the PCI condition was

compared with the CPI condition. Thus, path learningpreceded conjunctive learning for some participants but notfor others. This comparison provided evidence for contrast,with more conjunctive choices (and thus less attention tointrinsic motion) in the CPI condition than in the PCIcondition, t(92) = 2.42, p < .025. T his finding suggests thatprior path learning resulted in greater attention to intrinsicmotion during conjunctive learning, helping participants inthe PCI condition to distinguish conjunctive verbs from pathverbs.A second comparison tested for persistence to intrinsicmotion during conjunctive learning as a result of prior

intrinsic motion learning. In particular, the PIC conditionwas compared with the PCI condition. Thus, intrinsicmotion learning preceded conjunctive learning for someparticipants but not for others. This comparison providedevidence for persistence, with more conjunctive choices(and thus less attention to intrinsic motion) in the PCIcondition than in the PIC condition, /(92) = 2.65, p < .01.This finding suggests that prior intrinsic motion learningresulted in greater attention to intrinsic motion duringconjunctive learning for participants in the PIC condition.Given a fixed total pool of attentional resources, increasesin attention to intrinsic motion would be accompanied bydecreases in attention to path. The results of the intrinsicmotion trials can be used as a measure of attention to pathduring conjunctive learning. If participants noticed the pathcomponent of conjunctive verbs, they should have rejectedconjunctive verbs as labels for the intrinsic motion trials.Increases in attention to intrinsic motion as a result ofpersistence or contrast, however, may have resulted indecreases in attention to path, making it more difficult for

participants to distinguish conjunctive verbs from intrinsicverbs.The percentages of conjunctive choices in the intrinsicmotion trials are depicted in Figure 9. An ANOVA on thesepercentages revealed a main effect of learning order, F(2,89) = 3.23, p < .05, MSE = 573.36. There was no maineffect of the choice of intrinsic motion attribute, F(\, 89) =1.53, p > .10, MSE = 573.36, nor did this variable interactwith learning order, F( l, 89) < 1. Two nonorthogonalplanned comparisons were carried out to understand themain effect of learning order on the percentages of conjunc-tive choices in the intrinsic motion trials. One comparison


15/22

COMPETING ATTENTIONAL MECHANISMS 1451

P -> I -> CP -> C -> IC -> P -> I

Ma nne r Leg Mot ionIn t r i ns i c Mot i on A t t r i bu teFigure 9. The percentages of conjunctive choices in the intrinsic motion final test trials ofExperiment 2 that matched the presented event on its path component but mismatched on path. Errorbars reflect standard errors. PIC = path intrinsic motion (i.e., manner of motion or leg m otion) conjunctive; PCI = path conjunctive intrinsic motion; CPI = conjunctive path * intrinsicmotion.

examined whether persistence to leg motion resulted in adecrease in attention to path. This comparison compared thePIC and PCI conditions. Thus, intrinsic motion learningpreceded conjunctive learning for some participants but notfor others. The results of this analysis approached signifi-cance, r(92) = 1.84, p < .07, with a trend toward moreconjunctive choices (and thus less attention to path) in thePIC condition than in the PCI condition. Thu s, prior intrinsicmotion learning may have resulted in less attention to pathduring conjunctive learning for participants in the PICcondition.

A second comparison examined whether increases inattention to intrinsic motion because of contrast resulted indecreases in attention to path. This comparison comparedthe PCI and CPI conditions. Thus, path learning precededconjunctive learning for som e participants but not for others.This co mparison was not significant, t(92) = 0.65, p > .10.Thus, although there is evidence that prior path learningresulted in greater attention to intrinsic motion because ofcontrast, Experiment 2 does not provide evidence that thisprior learning resulted in decreases in attention to path.

DiscussionAs in Experiment 1, participants were better able todistinguish path and conjunctive verbs when conjunctiveverbs were learned after path or intrinsic motion verbs.Participants who learned path verbs before conjunctiveverbs made fewer conjunctive responses in the path trialsthan did participants who learned conjunctive verbs first.This provides evidence for contrast in that participants w eremore likely to notice the intrinsic motion component of

conjunctive verbs as a result of prior path learning. Partici-pants who learned intrinsic motion verbs before conjunctiveverbs made fewer incorrect conjunctive responses in thepath trials than did participants who learned conjunctiveverbs before intrinsic motion verbs. This provides evidencefor persistence, in that participants were more likely tonotice the intrinsic motion component of conjunctive verbsas a result of prior intrinsic motion learning.Experiment 2 also provided evidence that increases inattention to leg motion because of persistence resulted indecreases in attention to p ath. Experiment 2 did not provideevidence, however, that increases in attention to leg motionas a result of contrast are also associated with decreases inattention to path. It is possible that a floor effect wasresponsible for this lack of effect. Experiment 3 moresystematically tested whether increases in attention to oneattribute because of persistence and contrast result indecreases in attention to other attributes.

Exper imen t 3Experiment 3 was designed to test the hypothesis thatpersistence operates primarily at the level of attributes,whereas contrast operates at the level of attribute va lues. Toprovide evidence for persistence to attributes, we compareda control condition in which conjunctive verbs were learnedfirst with two conditions in which leg motion verbs werelearned prior to conjunctive verbs (see Table 3). If persis-tence operates at the level of attributes, then prior learning ofleg motion verbs should result in greater attention to legmotion during subsequent conjunctive learning. As a result,participants in the control condition w ere expected to attend


16/22

1452 KERSTEN, GOLDSTONE, AND SCHAFFERTTable 3Schematic Description of the Three Learning Conditions in Experiment 3

ConditionControlPersistenceContrast

Phase 1 categoriesPT Li vs . P2 L2(Conjunctive)P 12 L 3 vs . P,2 L 4(Different leg motion)P12L1 vs.P| 2 L 2(Identical leg motion)

Phase 2 categoriesP 1 2 L, VS.P12L2(Identical leg motion)PiLiVs.P2L2(Conjunctive)Pi Li vs. P 2 L2(Conjunctive)

Phase 3 categoriesP12 L 3 v s . P j 2 L4(Different leg motion)P12L1 vs. P 1 2 L 2(Identical leg motion)Pl2 U VS. Pj2 U(Different leg motion)

Note. Subscript numbers represent values of leg motion and path. A subscript 12 following anattribute indicates that the attribute was nondiagnostic, taking each value in half of the examples of averb. Each phase of learning is labeled by parenthetical text appearing beneath the schema for thatphase. P = path; L = leg motion.

less to leg motion during conjunctive learning than wereparticipants in the other two conditions.

To provide evidence for contrast to attribute values, thetwo conditions involving prior leg motion verb learningvaried in the particular values of leg motion that were used.In the persistence condition, different values of leg motionwere involved in conjunctive verb learning and prior legmotion learning. Because the values of leg motion displayedduring conjunctive learning had not been encountered beforein this condition, no contrast effects were expected tocounteract effects of persistence to leg motion as an at-tribute. In the contrast condition, the same values of legmotion were involved in conjunctive verb learning and priorleg motion verb learning. Because the two values of legmotion displayed during conjunctive learning would alreadyhave been associated with categories, contrast effects wereexpected to partially offset any effects of persistence in thiscondition. Thus, participants were expected to attend less toleg motion in the contrast condition than in the persistencecondition.

MethodParticipants

One hundred sixty-eight undergraduates at Indiana Universitytook part in this experiment in partial fulfillment of courserequirements for introductory psychology.Stimuli

Learning events. There were three phases of learning. In theconjunctive phase, each of the two verbs labeled a particularcombination of path and leg motion. In the identical leg motionphase, each of the two verbs labeled a leg motion that was identicalto the leg motions presented in the conjunctive phase. In thedifferent leg motion phase, each of the two verbs labeled a legmotion different from the two that were presented in the conjunc-tive phase. Path varied orthogonally to verb meaning in both legmotion phases of learning- Thus, a verb from one of the leg motionphases of learning could be accompanied by either of the two pathsbut was always accompanied by the same leg motion.Participants in the control condition received the conjunctivephase first, followed by the identical leg motion phase and then thedifferent leg mo tion phase. Participants in the persistence cond itionreceived the different leg motion phase first, followed by the

conjunctive phase and then the identical leg motion phase.Participants in the contrast condition received the identical legmotion phase first, followed by the conjunctive phase, and then thedifferent leg motion phase.Criterion test trials. These w ere generated in the same way asthose in Experiment 2.Final test events. These were different from those of Experi-ment 2 in that only two verbs were presented as possible labels foreach event. There were seven different test types. Table 4 shows aschematic description of the final test types. Three tested forattention to leg motion. In the first type (Leg Motion 1), no pathwas show n. The agent sat in place moving its legs in a particular legmotion. Participants chose between the two conjunctive verbs, oneconsistent with the presented leg motion and one inconsistent. Inthe second type (Leg Motion 2), a path and leg motion from theconjunctive phase were presented butwere paired differently thanin the conjunctive phase. Participants chose between a conjunctiveverb consistent with the presented path but inconsistent with thepresented leg motion and a verb from the identical leg motionphase that was consistent with both the leg motion and path.Making this second choice would indicate attention to leg motion.In the third type (Leg Motion 3), a path from the conjunctive phaseand a leg motion from the different leg motion phase werepresented. Participants chose between a verb from the conjunctivephase consistent with the path but inconsistent with the leg motionan d a verb from the different leg motion phase consiste nt with bothleg motion and path. Making this second choice would indicateattention to leg m otion.

Four test types measured attention to path. In the first type (Path1), no leg motion was shown. The agent m oved along its path whileholding its legs stationary. Participants chose between two conjunc-tive verbs, one consistent with the presented path and oneinconsistent. In the second type (Path 2), a path and leg motionfrom the conjunctive phase were presented but were paireddifferently than in the conjunctive phase. Participants chosebetween a conjunctive verb consistent with leg motion but inconsis-tent with path and a leg motion verb that was consistent with bothpath and leg motion. Making this second choice would indicateattention to path. In the third type (Path 3), path was notrepresented. Participants chose between a conjunctive verb thatwas consistent with the presented leg motion but also madepredictions forpath and a leg motion verb that was consistent withthe presented leg motion; participants m ade no predictions for path.Making this second choice would indicate attention to path. In thefourth type (Path 4), a path and leg motion from the conjunctivephase were presented, paired in the same way as in the conjunctivephase. Participants chose between a conjunctive verb and anidentical leg motion verb, both of which were consistent with both


17/22

COMPETING ATTENTIONAL MECHANISMS 1453Table 4Schematic Description of the Test Stimuli in Experiment 3

Learning stimulusP J L JPi 2L,P12UFn U

Verbmorpingwunkingspoggingzellingyimminggupping

Test typeLeg Motion 1Leg Motion 1Leg Motion 2Leg Motion 2Leg Motion 3Leg Motion 3Leg M otion 3Leg Motion 3P a t h lP a t h lPath 2Path 2Path 3Path 3Path 4Path 4

StimulusPoLiP0L2PiLjPiL,P1L3P1L4P2L3P1L4P1L0P2L0P , L 2P 2 L ,PoL,P0L2P,L,P2L2

Verb choicemorping or wunkingmorping or wunkingmorping or zellingwunking or spoggingmorping or yimmingmorping or guppingwunking or yimmingwunking or guppingmorping or wunkingmorping or wunkingwunking or zellingmorping or spoggingmorping or spoggingwunking or zellingmorping or spoggingwunking or zelling

Note. Subscript numbers represent values of leg motion and path. A value of 0 indicates that thevalue of that attribute was not represented in the stimuli. P = path; L = leg motion.

path and leg motion. The conjunctive verb specifically predictedthe presented path, however, whereas the leg motion verb made nopredictions for path. Thus, participants were expected to choose theconjunctive verb to the extent that they attended to the pathcomponents of these verbs.

variable for these analyses was learning order (control v s. persis-tence vs. contrast). A second dependent variable was performancein the criterion test trials. The independent variables for thisanalysis were learning order and the type of verb being tested(conjunctive vs. different leg motion vs. identical leg motion).Procedure

The procedure was identical to that of Experiment 2 except thatthere were 32 final test trials. Each of th e 16 test types listed inTable 4 was presented twice.

DesignThe primary dependent m easures were the percentages of correctresponses to the 7 different final test types. The independent


The results of the first block of criterion test trials witheach verb type are displayed in Figure 10 . As in Experiments1 and 2, the relative difficulty of the different verb typesprovides a measure of attention to path and leg motion.These results were analyzed in an ANOVA with learningorder and verb type as independent variables. This analysis

10 0Z 90< 5g 70oQ >Q- 60

501it 11

T

*'/,',\

ControlPersistenceContrast

Conj u nct iveD if f e rent Ident ica lVerb Type

Figure 10. Resu lts of the first block of criterion test trials for each verb type in Experiment 3 . Errorbars reflect standard errors.


18/22

1454 KERSTEN, GOLDSTONE, AND SCHAFFERTrevealed no main effect of learning order, F( 2,1 65 ) < 1. Theeffect of verb type w as analyzed using two planned co mpari-sons. Thefirstcomparison com pared the learning of conjunc-tive verbs with the learning of the two types of leg motionverbs. Conjunctive verbs were expected to be easier b ecausethese verbs could be differentiated on the basis of path. Thiscomparison revealed better performance with conjunctiveverbs than with the two types of leg motion verbs, F(l,165) = 14.96, p < .001, M SE = 406.17. This comparisonwas not dependent on learning order, as indicated by theabsence of an interaction, F(2 ,16 5) < 1.

The second comparison compared the learning of identi-cal leg motion verbs with the learning of different leg motionverbs. This com parison revealed better learning of differentleg motion verbs, F{\, 165) = 13.69, p < .001, MSE =647.08. This comparison was also dependent on learningorder, as indicated by a significant interaction, F(2, 165) =13.76, p < .001, MSE = 647.08. Post hoc LSD tests wereconducted separately on the two types of leg motion verbs tounderstand these effects. An analysis of identical leg m otionverbs revealed that participants in the persistence conditionperformed better than participants in the other two condi-tions, which did not differ. T his advantage of the persistencecondition m ay reflect persistence to leg m otion as a result ofprior learning of different leg motion verbs. Participants inthe other two conditions learned identical leg motion verbsbefore different leg motion ve rbs. An ana lysis of differentleg motion verbs revealed that participants in the persistencecondition performed worse than did participants in the othertwo conditions, which did not differ. This advantage of thecontrol and contrast conditions is also consistent withpersistence to leg motion. Participants in these conditionslearned identical leg motion verbs prior to different legmotion ve rbs, whereas participants in the persistence cond i-tion learned different leg motion verbs first.

Final Test TrialsAs in Experiments 1 and 2, participants were selected foranalysis of final test scores if they performed with 100%accuracy on at least one block of criterion test trials for eachverb type. Out of 56 participants in each condition, thenumbers of participants in each condition who achieved this

criterion were as follows: 43 in the control condition, 50 inthe persistence condition, and 46 in the contrast condition.The results of the three different tests of attention to legmotion are depicted in Figure 11. Th e effects of learnin gorder on performance in the leg motion trials were analyzedusing two orthogonal planned comparisons. The first com-parison tested for attentional persistence. This comparisoncompared the control condition with the persistence andcontrast conditions. These last two conditions w ere groupedtogether in this analysis because they both had a leg motionphase of learning prior to conjunctive learning. This wouldbe expected to draw attention to leg motion as an attribute.Analyzing Leg Motion 1 trials in this way revealed signifi-cant persistence, /(136) = 3Al,p < .001, with less attentionto leg motion in the control condition compared with theother two co nditions. The analysis of Leg Motion 2 trials didnot reveal significant persistence, although the trend w as inthe predicted direction, f(136) = 0.80, p > .10. Leg Motion3 trials also failed to reveal significant persistence. This lackof effect was apparently a result of a ceiling effect, withnearly perfect performance in all three conditions.The second planned comparison tested for contrast. Thiscomparison compared the persistence condition with thecontrast condition. These two conditions varied in theparticular values of leg motion that were used. Participantsin the contrast condition saw the same values of leg motionin the first phase of learning and in subsequent conjunctive

learning. Participants in the persistence condition sawdifferent values of leg motion in the first phase of learning

1 U U "

o 9 0 "CD5 80-o*- 70-cCDif 6 0 -oQ . 5 0 -

^n -

-r TT "T-r a t e "1II1T l H PHIHHJ

T T i1 - I - J' v

S \

\ \

' v\ s\ S

\ \/ \ \

*

1

d1

mmTTi b

1mma

. . . 'v -/ sV,/ / / /

S \y

\ \ '/ X ^


19/22

COMPETING ATTENTIONAL MECHANISMS 1455and in subsequent conjunctive learning. Thus, contrasteffects were expected in the contrast condition but not in thepersistence condition. Analyzing the Leg Motion 1 trials inthis way revealed evidence for contrast, f(136) = 2.04, p .10, although the trend was in the predicted direction.There was a significant effect of contrast in the Path 2 trials,f(136) = 1.98, p < .05, with greater attention to path in thecontrast condition than in the persistence condition. Therewere no significant effects of contrast in the Path 3 or Path 4trial.

DiscussionExperiment 3 provides evidence that attentional persis-tence operates at the level of attributes, whereas contrasteffects are realized at the level of attribute values. Inparticular, persistence following leg motion training resultedin increased attention to leg motion and decreased attention

to path, compared with a control condition that was given n oprior leg motion training. If the same values of leg motionwere seen in the first two phases of learning, however,contrast resulted in decreased attention to leg motion andincreased attention to path, compared with a condition inwhich different values of leg motion were seen in the firsttwo phases.Although the results of Experiment 3 are generally quiteconsistent with both persistence and contrast, not every testtype revealed evidence for both attentional mecha nisms. Thenull results of Leg Motion 3 and Path 4 trials can be quiteeasily explained in terms of ceiling and floor effects,respectively. Even the rem aining test types, however, differin the extent to which they reveal persistence and contrast.For example, the Leg Mo tion 1 trials reveal a greater am ountof persistence than do the Leg Motion 2 trials, even thoughthe two types of trials are conceptually similar. In addition,the Path 2 trials reveal a greater amount of contrast than dothe Path 1 trials, whereas the Path 3 trials did not reveal evena trend toward contrast. At present, it is unclear why themagnitude of persistence and contrast effects differed acrossthese different test types.The results of Experiment 3 are for the most partconsistent with a model in which the total pool of attention isfixed (e.g., Nosofsky, 1986; Sutherland & Mackintosh,1971). In such a m odel, increases in attention to one attributeresult in decreased attention to other attributes. In Experi-ment 3, increases in attention to leg motion were apparentlyaccompanied by decreases in attention to path. The most

ControlPersistenceContrast

2 3 4Path Test Type

Figure 12. Results of the path final test trials in Experim ent 3. Error bars reflect standard errors .


20/22

1456 KERSTEN, GOLDSTONE, AND SCHAFFERTstraightforward comparison of attention to leg motion andpath involves the Leg Motion 1 and Path 1 trials, which werehighly analogous. Such a comparison revealed that the totalamount of attention to leg motion and path was quite similaracross the three conditions, even though the distribution ofattention over these two attributes varied considerably. Thisnull finding, however, does not disconfima models that allowchanges in the total pool of attention (e.g., Mackintosh,1975).

General Discuss ionThe present experiments provide evidence for two compet-ing attentional mechanisms. Attentional persistence servesto increase attention to attributes previously found to bepredictive. Contrast serves to increase attention to attributevalues whose predictive power has yet to be discovered. InExperiment 1, we demonstrated that attention to manner ofmotion, an attribute w hose salience w as initially low, couldbe increased as a result of prior training in which mann er of

motion was relevant. This finding provides evidence forattentional persistence. Attention to manner was also in-creased following path training, providing evidence forcontrast. The interpretation of contrast effects in Experiment1, however, was complicated by the fact that manner ofmotion did not vary when only path was relevant. Thus,increased attention to manner following path learning couldhave simply reflected attention to novel variation on manner.Experiment 2 provided evidence for contrast by using adesign in which no new variation on an attribute wasintroduced. Experiment 3 provided evidence that persistenceoperates primarily at the level of attributes, whereas contrastoperates primarily at the level of attribute values.

Feature Creation and Attentional MechanismsAlthough the present task is similar to the task Schyns andRodet (1997) used to provide evidence for feature creation,the current results can be explained entirely in terms ofchanges in attention weights. Feature creation and atten-tional mechanisms may, however, play intimately relatedroles in categorization. When confronted with a novelcategory label, a categorizer may look first to his or herexisting featural repertoire to determine whether any ofthose features can be found in the sensory environment.Attentional persistence may encourage the sampling of

features from some attributes over others. At the same time,contrast may direct attention to features that have notalready been associated with a category. If there are noexisting features that match the present situation and are notassociated with a category, contrast may encourage thecreation of a new psychological feature based on somesalient physical property of the present environment. Atten-tional persistence may again play a role in this process offeature creation, determining which physical properties aremost salient. For example, after learning a num ber of wordsthat refer to shape, a child may be more likely to create anew shape feature to categorize a novel object. An associa-

tion can then be created between this new feature and itscorresponding category.The contrast mechanism proposed in the present articleassumes the presence of features that have not already beenassociated with a category. If new features are only createdwhen they are necessary to distinguish a new category fromexisting categories, then the present contrast mechanism isidentical to the feature creation mechanism proposed bySchyns and Rodet (1997). If this were the case, all featuresin a learner's featural repertoir

kersten 1998 two competing attentional mechanisms in category learning

Documents