Relational Similarity and the Similarity Judgments similarity judg- . - PDF document

Relational Similarity and the Similarity Judgments similarity judg- these theories have explanatory power domains partially role of relational judgments. Previous structural and aspects of notion of relational similarity by the fact that and her some ways there is in which probably not share some shared by account of two scenes in both incorporating relations way that and the abstract feature alternative hypothesis cannot adequately influence similarity factors that model of which specifically which weights the two features present separate features includes relational important generalizations concerning thought of write “GREEN(GRASS)” takes the “pretty” could attributes in tation system. description of attributes. Relations predicates which relations can the picture has the relation “Same shading.” propositional representation is often (CIRCLE, TRIANGLE).,, The is critical relation, ordered While “ABOVE CIRCLE)” describes a a circle, TRIANGLE)” describes above triangle. to the “RIGHT-OF(TRIANGLE, CIRCLE),” attributionally similar although the relation between the differs. Attributional similarity and relational similarity more global than sense that general, however, relations cannot the same face) and could not single relation. the distinction between rela- and global this paper the representation system. attributes; another system, lines a series “Adjacent-to” relations. lines. “Parallel possible relation property is an once a particular representation commanding influence and mathematical equa- function of features that features that has that minus again terms that depend is not inherent requirement additivity such three assumptions: states that function of their shared features features. This from analysis features that number of shared independence assumption independent of third component. is incorrect independence; there three components that enter similarity judgment: of features, without having independence when itative and quantitative same stimulus feature causes but not shared feature computed by feature (closure) that not. Consequently, Gati and that feature violated when with old produce emergent features. that is more of squares. figure with have reversed original similarity interaction between and the top each is in separately judgment reversal, between features framework of could simply that relations should In the have considered features, relations in Tversky's and these judgments. In have suggested that in Fig. have the of relations would require features" hypothesis distinct ways. show violations of 227 RELATIONAL SIMILARITY element of element. Exactly same feature change occurs between differences between similarity ratings we can that the difference between how similar and how result required similarity models feature change second expression independence assumption is increases rated stimuli than results, stating final similarity the size of the get more tributes, then attributes functional differentiation claim selective attention claim. distinct, that type is third expression would independence assumption, attributional similarity it is relational similarities This is the model, which similarities belonging selectively weighted models is RELATIONAL SIMILARITY relations more; doesn't depend important points about that they they both require similarity types: is that attributional feature. that treats relations model which selectively weights relations context-independent manner objects; each varying on shading. shape, group containing Tpicture for the groups were larity with relational similarity relational similarities include Taller-than, Larger-than, similarities include the question second test were arranged were designcd printed on the handout and were the choice which the comparison figure. The took approximately pictures (the first six above) were (the last forced choice measure of relative similarity relative similarity conservative measure subjects who in accor- the number independence model is supported with “independent” group contain responses based Instead, systematic from independence measured by comparing “noise” responses could responses should equal for independence assumption. all nine picture times, MIN effect were observed independence group. greater than numbers supporting number of by picture relational similarities attributional similarities. of the responses that relational similarities preferring attributional relational similarities over the model suggests that attributional similarities important for judgments when two objects being compared (and conversely relational similarities). relationally similar choice, tional similarity attributional choice, was picked when the overall attributional similarity particularly strik- the within-subject RELATIONAL SIMILARITY plausible possibility choices map ratio of rather than contribution of Referring again attributional matches, with this greater choices than advantage of when total ratio rule interpretation by varying relational similarities. is that either down relational similarity. more total than attributes, interpretation of nonindependence of course, predict that the choices similarities, respectively. ratings from Experiment each element varying size, shading, shape. Each group a single four members attributional similarity the size, any of their elements. common with and one more Same-shading-as, Taller-than, Larger-than, relational similarities and one six different comparisons. slot was comparisons from Participants were set. whichever took approximately taking part in Results and Discussion ordinal measure conservative measure independence assumption less than second term. ordinal results Otherwise: Independence. must occur. then clearly A, then pattern requires same feature which stimulus subjects, and eight pictures, with independence. of the independent group, this group cannot compared with greater than relational picture choose between is the Pictures with relational similarities generally judged similarities (Relational attributional pictures ratio rule interpretation nonindependence must which claims, weighted.” While this hypothesis attributional choices more common among respond according subject. Seventeen subjects made more made more three subjects biased and subject failed One objection that the feature that feature that the feature Shading symmetry” going from is added. for the stimuli in picture groups not the additional fea- The four that were possibly manifesting higher- four asterisked items were greater than weight of in pictures already relationally similar. already attributionally similar. the relative attribute “BLACK.” attributionally alike is more relationally alike important than The second experiment undermines n does the relation the attrib- demonstrated ordinal independence that senting choices context sensitivity concern that are a forced choice task for the Overcome by forms. The ratings follow change from relational similarity similarity ratings, of attribu- between the monotonically increasing, this pattern for the were six pictures fitting up of between two and relations used were disrupt the other pictures were dis- They were told pictures were keys for intermediary degrees picture and its group were displayed them some idea variability. During period subjects Next, the subjects and the picture was ized, and counterbalanced second display the subject. Consequently, each target picture was the four Subjects were tested simultaneously. similarity ratings the four picture is greater that the hand, predicts and an six picture is greater than zero. over the picture groups, comparable magnitude conditions prevented 4 5 GOLDSTONE, MEDIN, butional similarities. crossover occurs is more similar more similar found for three of the picture similarity ratings for picture set (nonsignificantly) than fact that similarity ratings is suggestive effect that each picture more similar pattern applied. the subject’s same ordinal counted as therefore left unlabeled, this analysis, except that given, the similar the objects are. number of greater than the number subjects for and the whether an relation match has that the pitted against similarity more the objects. that this interval criteria. was designed elements between picture sets. the objection claim would add the have the identical In going match without adding IDENTICAL OBJECT along all significance above represent the has recently adults do give par- identicality. This objection stimuli like on similarity we may second objection tached parts constitute the shapes are detached that one scene does touch any scene does Gestalt (Wertheimer, that is, a coherent, met by designing stimuli that never advantage to to A, B, C, Dl picture group. An example of one such set is shown in Fig. 5. In this set, going from T - C to T - A does not result in an identical part match because none other sets, parts are second objection is perceived are compared identical except that the large the scenes Fourteen picture that subscribed were never exactly identical insure nonidenti- the appcarancc shading, size, parentheses show the mouth. similarity ratings for each subjects. For is positive, interval criteria for picture set The overall the difference sets, there ordinal crossovers direction. In these gets a higher rating Similarity Ratings 3.4 3.2 ~~ ~~~ 6.3 6.3 circko repnscnting control, faceless each subject's response pattern sensitivity in Experiment two judgments in Table "No. MAX" subjects in Table except that paired with the second considered separately. labels were more similar was greater than the We can subject depending of their fourteen subject had exactly identical of MAX MIN responses. that it dismisses few strongly relational weighting during times large majority of people, strong, ordinal evidence of the two There is effects than difference between sets and control sets. perceptually detached. In addi- that attributes count more attributionally similar support the relative impor- attributes in object the types already present predominate, then attributional have their The bias :cts, Exper- fects for ed. In increased, relative relational matches. attributional similar- decreased (Experiments This pattern similarities between different effects similarity, depending features are similarity judg- similarity depends mechanisms function function even bias cannot explain attributional choices relationally, for independence assumption. THE OBSERVED the assumption of feature ditivity does served. Feature additivity the features that failure to found is given and shadings, also relations "Top and shapes are same." The taking Fig. an example, the contrast constrained to adopt the present data directly of Tversky's contrast embraces feature results do important processing principle that similarity judgment observed pattern processing assumption were included the other of Tversky’s is not sum of when different functional definition is underdetemined present experiments. larities of their own weights of the other termine overall already suggested relating similarity with such that of shared high [Hintzman’s cubic function, multiplicative rule, assumption of nonlinearity, that these using Medin multiplicative rule additive between- function, we is the the contrast attributes and distinct dimensions ordered (for three relations, two). MAX Tversky and argue for sional subadditivity by demonstrating of the that the different dimensions. these explanations of assumption that both corres and the RELATIONAL SIMILARITY values according SHAPE, and is even lack of of two domains where categories, not around effects circumvents with positing relational method requires crofeatures, coupled with two attributes features than attribute and do. Suppose number of features there both corresponding noncorresponding microfeatures. common features that match some of produce superadditivity any two chosen features dimension tend each other chosen features between dimensions, made that any crofeatures in As such, this account assumes that features must whether they variations have in common, attributes and similarity model divide features (at least) attributional types, relations between relations between features mutual influence arise when trying dimensional scaling model. multidimensional scaling points in the distance model (Carroll of stimulus the number of relevant and other demonstration in this MDS predict how make assumptions values that make it analysis cannot by ignoring dimensions. In general, such that the same cannot reverse surprising result judgments, it effects. In that assumes “relations” dimension (where two scenes dimension actually of this predict that because when pooled together their effects additive (unless similarities belong different dimensions. possessing differences be judged This, of exactly the opposite pattern between responses and those in tv a target. mapping. People about the goodness to selectively attend to observation that commonalities between that subjects into one such that classes. This make similarity the attributional from the and the this sorting the two similarity types can selectively distinction between relations and experiments, subjects Their subjects forced choice similarity judgments attributionally similar target and the target. They found relationally similar to the difference judgments appear to treat relations attributes differently; predict the differential effect on both is necessary whether the feature work is similar argued against the "segment-independence assumption" dence that than one argue against grounds that not have that the also stressed the features at for goodness, that attributes attributes are have major THE AlTRIBUTES/RELATIONS DISTINCTION drawing the tions is that distinctions. The these distinctions. such distinc- analytic (Brooks, these Jisinctions the relationslattributes distinction, that the former involves properties of the analytic see a that the paper has changed color also. the children perceived the difference old piece transformed piece, but were unable there is chy-if fe: of the is re the other to the that the relationally similar choice tended be chosen more similar the target and target. Similarity judgments appear attributes differently; to predict the differential effect necessary to the feature argued against the “segment-independence assumption” in light enter into larger more than argue against feature independence not have the other features that the objects possess. also stressed the three levels (whole figure, multisegment individual line postulated to divided, rated for goodness, influences on similarity THE ATTRIBUTES/RELATIONS the distinction and rela- is confounded other distinctions. that the relations these distinctions. such distinc- holistic vs analytic (Brooks, argue that these disinctions relations/attributes distinction, they neither nor jointly ex- haust its that the former the overall, latter apprehends the separate dimensions/ She gave evidence that its color also. that the children perceived the difference analyze the difference to determine the which was the change. to equate such the synthesis there is presents evidence that attributional interpretations “A cloud sponge,” the five year whereas the adult back later.” Thus, sensitivity to the relational bias, general attribute bias; pattern. Relative children, adults simultaneously more analytic (Kemler, sensitive to relations (Gentner, Further evidence against equating holistic comes from adult reaction Adults under pressure base judg- ments more properties than do adults 1984). On the other hand, our pressure increases the responses. In experiment, subjects ited time forced choice between required to respond within the attributionally more similar This percentage is significantly than the for subjects no time In other words, the more time to respond, the more likely the relationally over the attributional choice. quickly (developmentally hended, whereas to relations develops to the second distinction, can be processed without breaking object into parts (Navon, relates the construction hierar- more global Navon gives an example most global this scene is ‘blob” obtained celestial bodies. This can be be processed before resolved into objects. In object are processed before presented letters that example). Even were equally was influenced the global argument could including the involving the the relation only involves two out most global according to Navon’s precedence hypothesis even analyzed objects. The (second object, relation, however, order to see their components it would the actual shapes primitive symmetry detector. However, quite different components are. simple global detail. This since relations neither always nor perceptually their parts. from the stimuli is to define, the intuition attribute “gray.” conditions under abstract properties properties. Namely, two scenes superfrcially similar, further abstract similarity, then their superficial abstract correspondences. preferring the that its vest igat the design There are seems to characterize the differences the relation/attribute distinction cannot discussed above, procedures detailed investigate the dichotomy, the used. The large-scale similarities to small-scale similarity is judged similar to than is and would experimentation has and Palmer’s between global that the diverse distinctions psychological distinctions. shapes, and greater effect in the relational similarities equally plausible all similarities differentially specified attributes. Finding empir- evidence for deciding between separate structures same type are the relations pooled, and treated differently that these on the two such “fifth object striped” and shading” and “last object more than also sensitive . . ! 254 GOLDSTONE, MEDIN, AND GENTNER T m vlolmttons I mn vlolallons O T T nos vtolaltons 4 ntn v~o~e~ions I A C B D oc30 OAO A B OAO D oc30 C A B A B C D from the The choice-tz this model with the Additive generality, analyze on top these conventions, the figures choice-target similarities the feature additivity These results require is, the binary (circle it is first necessary relevant dimensions. Again general example, values of Bottom element, relation, taking values of W,, W,, have the have different these conventions (distance) expressions between between + w,*o + w,*v,c + w,*v,y d8,t = [WI*O + W,*V,' + W,*V{ + w,*v,y dc,t = [WI*VI' + w,*o + W,*V,' + w 41 *O I" dD,r = [Wl*Vl' + W2*V{ + W,*V{ + W,*O]"'. The systematic finding we reported was that subjects chose D over B, and A over C, as being most similar to T. Because the MDS model as- sumes similarity decreases as distance increases, our data However, it contradicts the previous requirement. picture group two groups: element: Circle or The second term is include two relations dimensions. change depending which they occur. The specific similarities added together. this metric, four choices The analysis differences between three choices. This inequality MEDIN, AND However, this MDS (plus (plus ()+ (Rl)?'" + [(Al)' + (R2)7'" requires that since (~2 + R1) � ((A2)' + (R1)7'" and (Al + R2) � [(AI)' + (R2)7'". Consequently, MDS predicts MIN, Psychology: Learning. und categorization. dimensional structure. three-way multidimensional C. Atkinson, and thought. Child Develop International Coderence Man, and cognitive development. cognition. and hierarchical patterns. relations: Judgments Psychology: General. perception and in cognition. Hierarchical structure visual similarity. Selective attention Representation of adult's classification: in children computing muchinery.