3

e

14

Rattermann, M. J ., & Gentner, D . (1998) . The effect of language on similarity : The use ofrelational labels improves young children's performance in a mapping task . In K. Holyoak, D. Gentner, & -B. Kokinov (Eds .), Advances in analogy research : Integration of theory and data from the cognitive,computational, and neural sciences (pp . 274-282). Sophia: New Bulgarian University .

The Effect of Language on Similarity :The Use of Relational Labels Improves

Young Children's Performance in a Mapping Task

Mary Jo Rattermann

Department of Psychology Whitely Psychology Labs Franklin and Marshall CollegeLancaster, PA 17604 M_Rattermann@ACAD.FANDM.EDU

Dedre GentnerDepartment of Psychology Northwestern University 2029 Sheridan Ave .

Evanston, IL 60208 gentner@nwu .edu

INTRODUCTION

were unable to preserve the plot strucwre intheir mapping, although they could transfer the

The ability to use relational similarity is

story plot accurately when given similar char-considered a hallmark of sophisticated think-

acters in similar roles. Older children (9-years-ing; it plays a role in theories of categoriza-

old) could maintain a focus on the relationaltion, inference, transfer of learning and gener-

structure and transfer the plot accurately despitealization (Gentner& Markman, 1997 ; Halford,

competing object matches . There is evidence1993 ; Holyoak & Thagard, 1995 ; Novick, 1988 ;

that this shift from objects to relations is basedRoss, 1989). However, young children often

on gains in knowledge (Brown, 1989 ; Goswa-fail to notice or use relational similarity (Gen- mi, 1993 ; Kotovsky & Gentner, 1996 ; Ratter-tier, 1988 ; Gentner & Rattermann, 1991 ; Gos-

mane & Gentner, in press), although ntatura-wami, 1993 ; Halford, 1993). For example,

tional changes may also play a role (Halford,when given the metaphor "plant stems are like Wilson, Guo, Gayler, Wiles & Stewart, 1995) .drinking straws" 5-year-old children focus on Children's ability to carry out purely rela-the common object similarities, commenting tional comparisons improves markedly acrossthat ' T ey are both long and thin," whereas 9- development . Yet even very young children canyear-olds focus on the relational commonality reason analogically under some circumstancesthat "They both carry water" (Gentner, !988) .

(Crisafi & Brown, 1986 ; Kotovsky and Gent-This relational shift in children's use of

ner, 1996). For example, Gentner (1977) dem-similarity-a shift from early attention to corn

onstrated that preschool children can performmon object properties to later attention to com-

a spatial analogy between the familiar basemon relational structure-has been noted

domain of the human body and simple picturedacross many different tasks and domains (Gen-

objects, such as trees and mountains . Whentner & Rattermann ; 1991 ; Halford, 1993) . For

asked, "if the tree had a knee, where would itinstance, Gentner and Toupin (1986) present-

be?," even 4-year-olds (as well as 6- and 8-year-ed children with a story mapping task in which

olds) were as accurate as adults in performingobject similarity and relational similarity were

the mapping of the human body to a picturedcross-mapped : that is, similar objects were

object, even when the orientation of the treeplaced in different relational roles in the two

was changed or when confusing surface at-scenarios, so that the plot-preserving relation-

tributes were added to the pictures .al correspondences were incompatible with

What factors impede or promok the per-obvious object-based correspondences . Under

ception of common relational structure? Ac-these conflict conditions, 6-year-old children

cording to structure-mapping theory (Gentner,

274

serve the plot structure in)ugh they could transfer they when given similar char-s. Older children (9-years-i a focus on the relationalr the plot accurately despitelatches. There is evidence,bjects to relations is basedlge (Brown, 1989 ; Goswa-& Gentner, 1996 ; Ratter-

1 press), although matura-also play a role (Halford,r, Wiles & Stewart, 1995) .ty to carry out purely rela-improves markedly acrossen very young children canunder some circumstances'986; Kotovsky and Gent-nple, Gentner (1977) dem-hool children can performetween the familiar basei body :and simple picturedes and mountains . Whenad a knee, where would itis (as well as 6- and 8-year-ite as adults in performinghuman body to a picturedhe orientation of the treeyen confusing surface at-to the pictures .ipede or promote the per-relational structure? Ac-mapping theory (Gentner,

rity:rovesAping Task

ad Marshall College)M.EDU

Sheridan Ave .

1983, 1989 ; Gentner & Markman, 1997) ananalogy is the mapping of knowledge from onedomain (the base) to another domain (the tar-get) in which the system of relations that holdsamong the base objects also holds among thetarget objects. When adults interpret an anal-ogy, the correspondences between base andtarget objects are based on common roles inthe matching relational structures ; the corre-sponding objects in the base and target do nothave to resemble each other . However, al-though the final interpretation of an analogyis determined by relational similarity ratherthan by object similarity, we hypothesize thatin the actual process of computing an analogyboth object similarity and relational similari-ty are at work (Falkenhainer, Forbus, & Gen-tner, 1990; Halford, Wilson, Guo, Gayler,Wile, & Stewart, 1995 ; Holyoak & Thagard,1989; Hummel & Holyoak, 1997 ; Keane &Brayshaw, 1988) .

A natural consequence of the structure-map-ping view is that knowledge of relations plays acrucial role in the mapping process ; if the child(or adult) has not represented the relations thathold within the domain then the matches formedwill be based upon common object similarityrather than common relational similarity . Thusas domain knowledge increases, so does the like-lihood that the child's comparisons will be basedon common relational structure .

In summary, the ability to use relationalsimilarity is sensitive to changes in the chil d'sknowledge base . With increasing knowledge ofthe relationships in a domain, children becomemore able to understand and produce purelyrelational matches. This brings us to the issueof interactions between language and thought .

Language and Relational Similarity

We propose that language may interactwith the development of analogical abilityby serving as an invitation to seek likeness-to make comparisons . The word-learningstudies of Markman, Waxman, and othershave shown that when children are taught anew object term they assume very that theword applies to things of like kind (Imai,

The Effect of Language onSimilarity

Gentner & Uchida, 1994 ; Markman, 1989 ;Waxman & Gelman, 1986) . However, thiswork has focused on noun learning . We pro-pose that the acquisition of relational lan-guage promotes the development of analo-gy by inviting children to notice and repre-sent higher-order relational structure (Gen-tner & Medina, 1997). So far, the evidenceon this issue is rather scant, although Ko-tovsky and Gentner (1996) found that 4-year-olds were better able to perceive cross-dimensional perceptual matches-e .g ., sym-metry of size compared to symmetry ofshading-when they had previously beentaught a relational label-"even"--to iden-tify the relation of symmetry .

The Present Studies

In these experiments we tested whetherchildren's relational performance can be im-proved by the introduction of relational labels .he basic task used in these experiments was a

cross-mapping search task in which object sim-ilarity and relational similarity were in conflictso that a response based on one type of similar-ity precluded a response based on the other. Wechose the higher-order relation of monotonicchange in size across position . This relation hasthe advantage that it can be understood on thebasis of perceptual information available to thechild (in contrast to some causal or social high-er-order relations that may require abstractknowledge). This cross-mapping task is usedin Experiment 1, whose results of serve as abaseline level of performance . In Experiment2, we gave children the relational labels "Dad-dy/Morn my/Baby" and found a predicted gainin performance . In Experiment 3 we tested othersets of relational labels, and further, tested forlong-term effects of labeling.

. EXPERIMENT 1

Participants

The participants were 24 3-year-olds, 244-year-olds , and 16 5-year-olds .

275

J

.e`

w

m

d

s

7

Procedure

Children were asked to map monotonicchange in size between a triad of objects belong-ing to the experimenter and a triad of objectsbelonging to the child . A cross-mapping was cre-ated by staggering the sizes of the objects, as il-lustrated in the following diagram in which theobjects, represented by numbers, form monoton-ic change in size from left to right .

E

3

2

1C

4

a

2The experimenter and the child sat across

from each other with the stimulus sets in twoarrays separated by about 6 inches, forming anarc in front of the child . The child closed hiseyes and the experimenter hid a sticker under-neath one of his toys, as she explained, "I'mgoing to hide my sticker underneath one of mytoys while you watch me . If you watch me care-fully, and think about where I hid my sticker,you'll be able to find your sticker underneathone of your toys." She then placed her stickerunder a toy in her set and said "If I put my stickerunder this toy, where do you think yours is?" .The child was then allowed to guess, but keptthe sticker only if he found it on his first guess .

Using a relative size rule, if the experiment-er chose object 2 in her set the correct choice isthe child's object 3 (Notice that the child mustresist an object match between the experiment-er's object 2 and his object 2 .) . Thus, object sim-ilarity was put in conflict with relational simi-larity (in the form of monotonic increase) form-ing a task in which a response based on eithersimilarity type is possible, but only a responsebased on relational similarity is correct . The chil-dren performed 14 cross-mapped trials .

Materials

We designed rich stimulus sets that containedinteresting, rich objects that varied along all di-mensions, including size, within the two sets (e .g .,a red flower in a pot, a wooden house, a greenmug, and a race car) and sparse stimulus sets thatcontained very simple, sparse objects that wereidentical in all respects but size within the twosets (e.g . . clay flower pots) . (See Figure 1 .).

276

Mary Jo Rattermann, Dedre Gentner

Based on structure-mapping theory, wepredicted that the rich object matches wouldcompete strongly with the relational mappingrule. In contrast, the sparse object matcheswould be relatively easy to overcome-childrenwould be able to perform the relational map-ping despite a common object identity choice .A related prediction was that the children wouldmake significantly more object-identity re-sponses when object richness was high thanwhen object richness was low .

Results and Discussion

The children's correct relational responsesrevealed both the predicted effect of object rich-ness and the relational shift in analogical per-formance . The richness effect led the childrento produce significantly more relational re-sponses with the sparse stimulus objects (54%c

for the 3-year-olds . 62% for the 4-year-olds, and95% for the 5-year-olds) than with the rich stim-ulus objects (32% for the 3-year-olds . 38% forthe 4-year-olds, and 68% for the 5-year-olds),suggesting that the presence of rich, distinctiveobject matches created a salient alternative tothe relational response (at least for young chil-dren) . In contrast . when sparse objects were.used, the object similarity matches were lesscompelling and therefore less likely to act as acompetitive alternative to the relational re-sponse. As further evidence for the effect ofobject richness, the number of object identityerrors significantly increased with the use ofthe rich stimuli (33% for the rich versus 17%

for the sparse . collapsed across all three agegroups). The relational shift was found in thechildren's overall performance, with the 5-year-olds performing significantly better than 3-and

Cross-Mapping Task

Sparse

Rich

E

CL

Figure 1.

ter

structure-mapping theory, wethe rich object matches would,ly with the relational mapping.st, the sparse object matchesely easy to overcome--childrento perform the relational map-ommon object identity choice .tion was that the children wouldntly more object-identity re->bject richness was high thanmess was low .

'Its and Discussion

)'s correct relational responsespredicted effect of object rich .,tional shift in analogical per-chness effect led the childrenifcantly more relational re-sparse stimulus objects (54%s. 62% for the 4-year-olds ; andr-olds) than with the rich stim-for the 3-year-olds, 38% fornd 68% for the 5-year-olds),presence of rich, distinctive

eated a salient alternative toonce (at least for young chil-, when sparse objects wereimilarity matches were lesserefore less likely to act as aiative to the relational re-- evidence for the effect of

number of object identityincreased with the use of

3% for the rich versus 17%apsed across all three age-)nal shift was found in the, rformance, with the 5-year-nificantly better than 3-ands-Mapping Task

Rich

i

4-year olds, who achieved above chance per-formance only with the sparse stimuli .

In Experiment 2 we tested whether a set ofrelational labels that provide children with anexplicit relational structure could help themcarry out a relational comparison and mapping .We introduced a group of 3-year-olds to the useof the labels "Daddy/Mommy/Baby" to de-scribe the relationship of monotonic change,and then presented them with the cross-map-ping task of Experiment 1 . We hypothesizedthat these labels would provide the children withan explicit framework for the relational systemofmonotonic change in size . If so, then the chil-dren's ability to perform a relational mappingwith both the rich and the sparse stimuli shouldimprove with the use of these labels . To obtainthe maximal effect of labeling, we went to greatlengths to ensure that the children were famil-iar with the relational use of the family labels,training participants with the "Daddy/Mommy/Baby" labels prior to presenting them with thecross-mapping task . We also reminded the chil-dren of these labels on each trial during thecourse of the experiment.

0 .2

The Effect of Language on Similarity

Rich

Stimulus Type

Sparse

EXPERIMENT 2

Participants

The participants were 24 3-year-olds.

Procedure

Label-training . The label training stimuliwere a set of toy penguins and a set of teddybears, each of four different sizes and with verydifferent markings. Training consisted of twosets of four trials in which the cross-mappingbetween objects and relations did not hold(bears were mapped to penguins) and two setsof four cross-mapped trials (penguins weremapped to penguins) . The following protocolwas used for the first eight trials :

"These bears and these penguins are eacha family. In your bear family, this is the Daddy(pointing to the larger bear) and this is the Mom-my (pointing to the smaller bear) . In my pen-guin family this is the Daddy and this is theMommy (again pointing appropriately) ." Whenthe child successfully labeled the animals in

I-

t'.

-1

I

A

I

w

.d

both sets, the experimenter said "If I put mysticker under my Daddy penguin, your stickeris under your Daddy bear . Look, my sticker isunder my Daddy . Where do you think yoursticker is?" and the child was allowed to searchfor the sticker, again only keeping it if he foundit on his first guess . After four trials a third,smaller, stuffed animal was added to each setand the labels "Daddy/Mommy/Baby" wereapplied in the manner described above . Thesame protocol was adapted for use in the cross-mapped penguin/penguin trials .

Cross-mapping trials . The cross-mappingtask from Experiment I was used. The childrenwere first asked to label both sets of objectsusing the family labels (this was repeated ev-ery second trial), and then the full-labeling pro-cedure (e .g ., "If I put my sticker under my dad-dy toy, your sticker is under your daddy toy .Look, my sticker is under my daddy, where doyou think your sticker is?") was used . The par-ticipants each performed 14 sparse trials and14 rich trials, counterbalanced, although onlytheir performance from the first stimuli typepresented was analyzed .

Results and Discussion

The use of the "DaddyA fommyBaby" la-bels did improve young children's ability tomake relational comparisons, even in the faceof a tempting object choice. When trained touse these labels, 3-year-old children's abilityto map relational similarity increased dramati-°cally with both rich and sparse stimulus sets .As can be seen in Figure 2,when "Daddy/Mom-my/Baby" was applied to the relation of mono-tonic increase, the number of relational respons-es produced by 3-year-olds increased from 54%with the rich stimuli and 32% with the sparsein Experiment I to 89% and 79%, respective-ly, bringing the performance of these partici-pants to the level of performance found in the5-year-olds . Note, however, that even when therelational labels were used, the effect of objectrichness was replicated ; children produced sig-nificantly more relational mappings with sparseobjects than with rich objects. Along with theincrease in relational mappings, there was a

278

Mary Jo Rattermann, Dedre Gentner

concomitant decrease in the number of objectidentity errors between Experiments I and 2(from 23% to 8% with sparse and from 43% to19% with rich) .

We propose that "Daddy/Mommy/Baby"helped the young children notice the presenceof a familiar higher-order relationship, namelymonotonic change, that they may have alreadyrepresented . Alternatively, the use of the rela-tional labels may have led children to align thetwo relational systems (the E set and the C set)and derive the common monotonicity structure .

In Experiment 3, we address three furtherissues . First, we asked whether relational ad-jectives such as "big/little/tiny" would also en-hance children's' ability to perform a relation-al mapping. Second, we tested for long-termrepresentational change brought about by ouruse of labels by retesting a sample of 3-year-olds 1-4 months after initial testing . And third,we addressed the possibility that our use of the"D-addy/MtommyBaby" labels on every trial inExperiment 2 led the children to use the labelsas an external crutch-perhaps following therule "look under the object to which the samelabel has been applied" without grasping therelationship of monotonic increase in size . Tobe able to dismiss this possibility we presentedchildren with a small number of full-label tri-als after which they were given new stimulussets and asked to perform the cross-mappingwithout labels being overtly applied .

EXPERLMENT 3

Participants

The participants were 51 3-year-olds, 1.8who returned to the laboratory for session 2 .The time period between Session I and Ses-sion 2 varied from I month to 4 months .

Materials and Procedure

Session 1 . The children were randomly as-signed to a labeling condition : no-labels, "Dad-dy/Mommy/Baby," or "big/little/tiny ." Chil-dren were given the label-training task used in

Experiment 2, and then eight t

the rich or the sparse stimuluslabel procedure of the previAfter completing the labeled t

were shown a new set of stirrichness type and were givenout labels .

Session 2. To ensure thatdon in this later session was assible from the initial sessionwere made; (1) the children vthe opposite type of stimuli (i .t

than was used in their initial(2) the children were tested ining room, and ; (3) a different eformed the experiment . The ir.to the children were minimal ; tted that they had played this gamember. you played a Dadd :game last time : Lets see if yougame." The children were gigtrials, without labels, using the .and bears. Each child was theneight unlabeled cross-mappingby four "reminder" trials in wbel procedure was used, . and teight more unlabeled trials .b

M

Results and Discus

Session 1 . Children trainetional labels ("Daddy/ommy/little/tiny'') produced significantal responses than children in thetion (58% for relational labelslabel, collapsed across stimulustrial type) . The effect of richneed; children produced signifcational mappings with the sparse sthe rich stimuli (67% versusacross label types and trial typeprevious experiments the childrenificantly more object identityrich than the sparse (37% versus

Session 2. We first examiability to map monotonic chateight, non-labeled, cross-rnappdata reflects children's abilityously leaned relational structure

cease in the number of object!! Between Experiments I and 2with sparse and from 43% to

that "Daddy/MOMMY/Baby"children notice the presence

er-order relationship, namelye, that they may have alreadymatively, the use of the rela-have led children to align theterns (the E set and the C set)rimon monotonicity structure .it 3, we address three furtherasked whether relational ad-)ig/little/tiny" would also en-ability to perform a relation-nd, we tested for long-termhange brought about by our:testing a sample of 3-year-Fter initial testing . And third,wssibility that our use of theBaby" labels on every trial in:he children to use the labelstch--perhaps following the'te object to which the samePlied" without grasping thenotonic increase in size . To:his possibility we presentedgall number of full-label td-v were given new stimulusperform the cross-mappingig overtly applied .

ERIh1ENT 3

irricipants

is were 51 3-year-olds, 28e laboratory for session 2 .nween Session I and Ses-I month to 4 months,

s and Procedure

children were randomly as-condition: no-labels, "Dad-oror "big/little/tiny ." Chil-label-training task used in

Experiment 2, and then eight trials using eitherthe rich or the sparse stimulus sets and the full-label procedure of the previous experiment .After completing the labeled trials the childrenwere shown a new set of stimuli of the samerichness type and were given eight trials with-out labels .

Session 2 . To ensure that the testing situa-tion in this later session was as different as pos-sible from the initial session several changeswere made; (1) the children were tested usingthe opposite type of stimuli (i .e., rich or sparse)than was used in their initial testing session ;(2) the children were tested in a different test-ing room, and ; (3) a different experimenter per-formed the experiment. The instructions givento the children were minimal, they were remind-ed that they had played this game before ; "Re-member, you played a Daddy/MommyBBabygame last time. Lets see if you can still play thegame." The children were given four practicetrials, without labels, using the stuffed penguinsand bears . Each child was then presented witheight unlabeled cross-mapping trials, followedby four "reminder" trials in which the full-la-bel procedure was used, and then finally witheight more unlabeled trials .

Results and DiscussionSession 1. Children trained with the rela-

tional labels ("Daddy/Mommy/Baby" and "big/little/tiny") produced significantly more relation-al responses than children in the no-label condi-tion (58% for relational labels and 41 % for no-label, collapsed across stimulus complexity andtrial type). The effect of richness was replicat-ed; children produced significantly more rela-tional mappings with the sparse stimuli than withthe rich stimuli (67% versus 39%, collapsedacross label types and trial type) . And as in theprevious experiments the children produced sig-nificantly more object identity errors with therich than the sparse (37% versus 20%) .

Session 2. We first examined children'sability to map monotonic change in the firsteight, non-labeled, cross-mapping trials . Thisdata reflects children's ability to apply previ-ously leaned relational structures with minimal

The Effect of Language on Similarity

prior reminding . The "Daddy/Mommy/Baby"and "big/little/tiny" labels led to more relationalresponses on these trials than did no labels (62%with the family labels and 54% with the rela-tional adjectives versus 28% with no-labels),suggesting that the children's previous expo-sure to relational labels had indeed changedtheir representation of monotonic change . Thesecond aspect of children's relational perfor-mance is their performance on the second setof non-labeled trials, after the four trial "re-minder" of the relational labels . Overall, chil-dren's relational responding increased afterbeing reminded of the relational labels (67%correct with relational labels, versus 45% cor-rect with no labels) .

We did not find a significant effect of objectrichness in this data due to the fact these childrenwere exposed to both types of stimuli across thetwo experimental sessions . It seems likely that thisexperience diluted the effect of object richnessfound in the previous experiments .

GENERAL DISCUSSION

A robust finding in the study of children'sanalogical abilities is the relational shift (Gent-ner, 1988; Gentner & Rattermann, 1991 ; Gentnerand Toupin, 1986; Halford, 1993) . In ExperimentI we explicitly tested for the relational shift andfound that the presence of a salient object simi-larity choice disrupted relational mapping in 3-and 4-year-old children, but that 5-year-olds couldmap relationally despite this conflict, supportingthe hypothesized shift from objects to relations inchildren's analogical reasoning .

In addition to testing for the relational shift,we also made predictions specific to the struc-ture-mapping view of analogy . The predictedeffect of object richness, one of the most robustfindings in this series of experiments, derivesdirectly from this view . We propose that whenperforming an analogical mapping, children (andadults) will begin by aligning objects based oncommon features, and further, that the more sa-lient and numerous the features, the more likelythat object matches will win out over relationalsimilarity in the final interpretations (Markman

27 9

1-1

A

& Gentner, 1993). In each of our experiments,the presence of a rich object conflict was moredetrimental to the ability to perform a relationalmapping than the presence of a sparse objectconflict. It is worth noting that a similar effecthas been found in the performance of adults pre-sented with a cross-mapping task . Markman andGentner (1993) found that adults will also re-spond based on object similarity when the num-ber of matching object attributes of the cross-mapped objects is increased .

In the present work young children's sus-ceptibility to rich object matches was due totheir incomplete knowledge of monotonicchange. We propose that simply using the la-bels "Daddy/Mommy/Baby," invited childrento represent the higher-order relation of mono-tonic change . We further claim that the addi-tion of this relational knowledge led to a strik-ing improvement-equivalent to that of a 2-year-age gain--in the children's ability to per-form relational mappings .

Finally, these experiments show quite force-fully that language, and in particular relationallanguage, can facilitate relational representation .We found that both "Daddy/Mommy/Baby" and"big/little/tiny" led to increased relational re-sponding in our three-year-olds, and that thisability remained several weeks after the initialexposure to these relational labels . The role oflanguage, we suggest, is to provide an invitationto form comparisons and further, to provide anindex for stable memory encoding of the newlyrepresented relational structure .

IMPLICATIONS FOR THEORIES OFANALOGY

The research of Halford and his colleagues(Halford, 1993 ; Halford, Smith, Dickson, May-bery, Kelly, Bain, & Stewart, 1995) has alsofound the shift from objects to relations. Theypropose that an important driver of this shift ischanges in cognitive capacity . That is, childrenshow a developmental increase in cognitivecapacity that allows them to represent and mapincreasingly more complex matches . Thus, forexample, not until three years should children

280

be able to carry out complex system matches.In contrast, in our account it is domain knowl-edge that leads to increases in children's ana-logical abilities .

Neither view of analogy is meant to be ex .clusive; we acknowledge the role of matura-tional change in children's cognitive abilitiesand Halford has consistently noted the role ofknowledge. However, our results demonstratethat striking changes in ability can occur overthe course of one experimental session, andfurther that these gains persist after the experi-mental session is over . It appears that the lim-its on performance are not in children's capac-ity to represent and use complex relations, butrather in whether they have as yet representeda given complex relation . These results under-score the point that an increase in relationalresponding is not evidence, in itself, of matu-rational gain .

Another prominent theory of analogicalreasoning is Goswami's (1993) relational pri-macy view . Goswami proposes that very youngchildren (3-years-old) can perform an analogywhen they have represented the requisite rela-tional structure . While we agree with 'oswa-mi that domain knowledge plays a crucial role,we differ in the hypothesized role of object sim-ilarity . Goswami has stated that "As long as therelations that the child must map can be repre-sented . . . . then performing the mapping shouldpresent little difficulty, and this should hold truewhether the objects to be mapped are similaror different in appearance." (Goswami, 1995,p. 891) . However, in our studies there is still arobust effect of object similarity ; even whenlabels have been applied and the children's re-lational performance is overall very good .

Language and Relations

We have presented the vieww that labels, andin particular relational labels, invite children tonotice and retain patterns of elements ; languageencourage them to modify thought. When ap-plied across a set of cases (or a pair of cases, ashere, Labels provide children with an invitationto make comparisons, and then provide a sys-tem of meanings upon which to base these com-

parisons (Gentner& Medina,of these labeling studies sup,year-olds trained with the labsmy/Baby" and "big/little/tin_nificant increase in relationala relatively simple linguistic

The results of these expthat young children can per -,mapping, even in the presenobject similarity, when familito highlight the appropriatecure. The impressive gains ituse of relational labels suppolanguage provides an invitaito modify their thought . Langever, the only path to relatioOther manipulations, suchalignmentin which children 1

easy literal similarity matchcult analogical matches willprovement (Kotovsky & Gentwith primates has also showlabels need not be embeddedtic system to improve relati,(Thompson, Oder & Boysen .

Thus we conclude that cdevelopment of the ability to u .,ilarity is the acquisition and ilanguage. Relational languagecatalyst for comparison and ajects and relations, which can,a mechanism for the progresdren's naive thought to the s'stract thought of adults .

REFERENCE:

Brown, A. L . (1989) . Analogi~transfer. What developsdou & A. Ortony (Eds .)analogical reasoningNew York: Cambridge L

Crisaft, M. A., & Brown . A . L.ical transfer in very y ,Combining two separat .lutions to reach a goal .went, 57, 953-968 .

jer

y out complex system matches .)ur account it is domain know]-to increases in children's ana-

w of analogy is meant to be ex-(nowledge the role of matura-n children's cognitive abilitiess consistently noted the role ofwever, our results demonstrateinges in ability can occur over,ne experimental session, ande gains persist after the experi .s over. It appears that the lim-ice are not in children's capac-and use complex relations, butr they have as yet representedrelation . These results under.that an increase in relationalt evidence, in itself, of matu-

minent theory of analogicalwami's (1993) relational pri-'ami proposes that very young-old) can perform an analogyepresented the requisite rela-Vhile we agree with Goswa-iowledge plays a crucial role,pothesized role of object sim-ias stated that "As long as the:hild must map can be repre

gforming the mapping shouldulty, and this should hold trueis to be mapped are similar)earance." (Goswami, 1995,in our studies there is still a:eject similarity, even when')plied and the children's re-ce is overall very good .

ge and Relations

ated the view that labels, andna! labels, invite children to'terns of elements ; languagernodifj' thought. When ap-cases (or a pair of cases, ccchildren with an invitations, and then provide a sys-!n which to base these com-

parisons (Gentner& Medina, 1997) .The resultsof these labeling studies support this view ; 3-year-olds trained with the labels "Daddy/Mom-my/Baby" and "big/little/tiny" showed a sig-nificant increase in relational responding witha relatively simple linguistic intervention .

The results of these experiments suggestthat young children can perform a relationalmapping, even in the presence of conflictingobject similarity, when familiar labels are usedto highlight the appropriate relational struc-ture . The impressive gains in ability after theuse of relational labels supports our claim thatlanguage provides an invitation for childrento modify their thought . Language is not, how-ever, the only path to relational competence .Other manipulations, such as progressivealignment in which children are presented witheasy literal similarity matches prior to diffi-cult analogical matches will also lead to im-provement (Kotovsky & Gentner, 1997) . Workwith primates has also shown that relationallabels need not be embedded in a full linguis-tic system to improve relational responding(Thompson, Oden & Boysen, 1997) .

Thus we conclude that one factor in thedevelopment ofthe ability to use relational sim-ilarity is the acquisition and use of relationallanguage . Relational language can serve as acatalyst for corrmpari~on and alignment of ob-jects and relations, which can, in turn, providea mechanism for the progression from chil-dren's naive thought to the sophisticated, ab-stract thought of adults .

REFERENCES

Brown, A . L . (1989) . Analogical learning andtransfer : What develops? In S . Vosnia-dou & A. Ortony (Eds.), Similarity andanalogical reasoning (pp . 369-412) .New York: Cambridge University Press .

Crisafi, M . A ., & Brown, A . L . (1986) . Analog-ical transfer in very young children :Combining two separately learned so-lutions to reach a goal . Child Develop-merit, 57, 953-968 .

The Effect of Language on Similarity

Falkenhainer, B ., Forbus, K . D ., & Gentner, D .(1990). The structure-mapping engine :Algorithm and examples . Artificial In-telligence, 41, 1-63 .

Gentner, D . (1977). Children's performance ona spatial analogies task . ChildDevelop-ment, 48,1034-1039.

Gentner, D. (1983) . Structure-mapping : A the-oretical framework for analogy . Cogni-tive Science, 7, 155-170.

Gentner, D . (1988). Metaphor as structure map-ping: The relational shift . Child Devel.-opment, 59,47-59 .

Gentner, D. (1989) . The mechanisms of ana-logical learning . In S . Vosniadou & A .Ortony (Eds .), Similarity and analogi-cal reasoning (pp . 199-241). London :Cambridge University Press .

Gentner, D. & Markman, A .G . (1997) . Struc-ture mapping in analogy and similarity .American Psychologist, 52, 45-56

Gentner, D. & Medina, J . (1997). Comparison andthe development of cognition, Cognition.

Gentner, D., & Rattermann, M . J . (1991) . Lan-guage and the career of similarity . In S . A .Gelman & . 3. P. Bymes (Eds.), Perspec-tives on thought and language- Interrela-tions in development . (pp . 225-277) . Lon-don: Cambridge University Press .

Gentner, D ., & Toupin, C . (1986) . Systematic-ity and surface similarity in the devel-opment of analogy . Cognitive Science,10, 27/7-300

Goswami, U. (1993). Analogical reasoning inchildren . Hillside, NJ : Lawrence Erlbaum .

Goswarni, U . (1995) . Transitive relational map-pings in three .- and four-year-olds : Theanalogy of Goldilocks and the ThreeBears Child Development, 66, 877-892 .

Halford, G. S . (1993) . Children's understand-ing: The development of mental models .Hillsdale, NJ: Erlbaum .

Halford, G. S . . Wilson, W. H ., Guo, J., Gayler,R . W ., Wiles, J., & Stewart, J . E. M .(1995). Connectionist implications forprocessing capacity limitations in anal-ogies . In K. J. Holyoak & J. Barnden

:''

•ks

282

(Eds .), Advances in connectionist andneural computation theory: Vol. 2. An-alogical connections (pp . 363-415) .Norwood, NJ: Ablex .

Holyoak, K . J ., & Thagard, P. R . (1995). Men-tal leaps : Analogy in creative thought.Cambridge, MA : MIT Press .

Hummel, J. E., & Holyoak, K. J. (1997). Dis-tributed representations of structure: Atheory of analogical access and mapping .Psychological Review, 104, 427-466 .

Imai, M., Gentner, D ., & Uchida, N. (1994) .Children's theories of word meaning : Therole of shape similarity in early acquisi-tion. Cognitive Development, 9, 45-75 .

Keane, M. T., & Brayshaw, M . (1988). The in-cremental analogical machine : A com-putational model of analogy . In D. SI-eeman (Ed .), Third European WorkingSession on Machine Learning (pp . 53-62). San Mateo, CA : Kaufmann .

Kotovsky, L, & Gentner, D . (1996) . Compari-son and categorization in the develop-ment of relational similarity . Child De-velopment, 67, 2797-2822 .

Markman, A . B ., & Gentner, D. (1993) . Structur-al alignment during similarity comparisons. Cognitive Psychology, 25,43 1-467 .

Mary Jo Rattermann, Dedre Gentner

Markman, E. M . (1989) . Categorization andnaming in children: Problems of induc .tion . Cambridge, MA : MIT Press

Novick, L. R . (1988) . Analogical transfer, prob-lem similarity, and expertise . Journal ofExperimental Psychology: Learning,Memory, and Cognition, 14, 510-520

Rattermann, M . J., & Gentner, D . (in press).More evidence for a relational shift inthe development of analogy : Children'sperformance on a causal-mapping task .Cognitive Development.

Ross, B . H . (1989) . Remindings in learning andinstruction . In S . Vosniadou & A. Ortony(Eds .), Similarity and analogical rea-soning (pp . 438-469). New York : Cam-bridge University Press .

Thompson, R . K . R ., Oden, D . L ., & Boysen, S .T. (1997) . Language-naive chimpanzees(Pan troglodytes) judge relations be-tween relations in a conceptual match-ing-to-sample task . Journal of Experi-mental Psychology: Animal BehaviorProcesses, 23, 31-43 .

Waxman, S . R., & Gelman, R . (1986) . Pre-schoolers' use of superordinate relationsin classification and language . CognitiveDevelopment, l, 139-156

ANALOGICA

I CREPCO U.R.A .2 Laboratoire "Psychol

U.P.R. de Ps

INTRODIJCT

Analogical reasoning itheme of the psychology ofment. It's probably because ivelopmemental mechanismsubject to construct and mocin a flexible and adaptive waOur goal is to analyse the coanalogical problem-solvingchildren .

Historically, studies oningg were first devoted to pgies (a :b : :c :d) then to sole(target problems) which reflerns (source problems) . in t

of view, the ability to reasones when the child reaches therecently, researchers showeddren did not fail because the\son by analogy but because tledge about objects or causal rmi &Brown, 1989 ; 1990 ar,terman, 1991). However mo :proposed to the children aregies which only allow to higlphase process . Other situationogles require to retrieve thesolve the target problem .

In problem solving situattons can contribute to impro