Full text of DOI:10.3758/BF03194966

Memory & Cognition
2002, 30 (4), 647-656
Dissociations between categorization
and similarity judgments as a result
of learning feature distributions
JEAN-PIERRE THIBAUT, MYRIAM DUPONT, and PATRICK ANSELME
University of Liège, Liège, Belgium
A dissociation between categorization and similarity was found by Rips (1989). In one experiment,
Rips found that a stimulus halfway between a pizza and a quarter was categorized as a pizza but was
rated as more similar to a quarter. Smith and Sloman (1994) discussed these results in terms of the role
of necessary and characteristic features. In two experiments, participants had to learn to categorize
novel artificialshapes composed of a nonsalient necessary feature combined with a salient character-
istic feature. Participants categorized stimuli on the basis of a necessary feature, whereas their simi-
larity judgments relied on characteristicfeatures. The role of deep (essential)features in dissociations
is considered. Results are discussed in terms of the differences between requirements of categoriza-
tion and similarity judgments.
According to many authors, similarity is a central con- estimate of a U.S. quarter size (1 in.) and theirsize estimate
cept for models of categorization in the sense that cate- of the smallest pizza (5 in.). The variance along the critical
gorization is grounded in similarity. Technically, an ob- dimension is different in the two categories. The size of a
ject X is categorizedin Category A instead of Category B quarter is fixed, whereas the size of pizzas is much more
if its representation is more similar to the representation variable. In the categorizationcondition,participantswere
of Category A than to the representation of Category B required to categorize the 3-in. target object in one of the
(see Goldstone, 1994; Hampton, 1998; Komatsu, 1992; two categories. In the similarity condition, participants
Murphy, 2002; Thibaut, 1997, for reviews of the relevant were asked to rate the similarity of the target with respect
literature). This similarity-based view is one of the dom- to the two categories. It was shown that whereas most
inant theories of categorization.
(63%) categorization participants categorized the target in
Contrary to this view, Rips (1989; see also Rips & the variable category (e.g., pizza), most similarity partici-
Collins, 1993) provided empirical evidence for a dissocia- pants(69%) found thetarget to be more similar to the fixed
tion between categorizationand similarity judgments. The category (e.g., U.S. quarters). This important result was
experimentalsetup consistedof a comparisonbetween cat- taken as evidence that categorizationis not based on simi-
egorizationand similarity judgmentsof a target stimulusX larity. In the case of categorization, most participants
with respect to two categories of stimuli, A and B. The ra- seemed to follow a rule.
tionale was that if the target was categorized in A more
Smith and Sloman (1994)tried to replicateRips’s (1989)
often than in B but judged more similar to B than to A, this results in two experiments. The instructions encouraged
result would demonstrate the dissociation. In one experi- participantsto use rule-based categorizationsby pointing
ment, Ripsread hisparticipantsa descriptionof a target ob- to the existence of a feature sufficient for categorization
ject described in terms of a value on a single dimension in one of the two categories (by “rule” we mean a set of
(e.g., the diameter). This value was chosen halfway be- properties—features—that, if they are satisfied by a stim-
tween the largest dimensionalvalue of a small category and ulus, will lead to its categorizationin a particularcategory).
the smallest value of a large category. To illustrate, a target In their first experiment, they did not replicate the disso-
3-in. objectwas chosen to be halfway between participants’ ciation obtained by Rips. They hypothesized that this
could be due to a procedural difference, in the sense that
in Rips’s experiment participants were instructed to talk
aloud as they made their choices, whereas Smith and Slo-
man’s participants were not asked to do so. Participants
Part of this research was presented at the Nineteenth Annual Meeting
of the Cognitive Science Society, August 1997 in Stanford and at the
were asked to do so in Smith and Sloman’s second exper-
iment. In the first experimental condition of this experi-
ment (the sparse condition),which was equivalentto Rips’s
experimental condition, choices in the similarity task
(50% of choices in favor of the variable category) signif-
icantly differed from the results in the categorizationtask
Simcat Conference, November 1997, in Edinburgh. The authors thank
Rob Goldstoneand PhilippeSchynsfor discussionson this research and
Gregory Murphyfor his generous comments on this manuscript. Steven
Sloman and two anonymous reviewers also contributed useful com-
ments. Correspondence should be addressed to J.-P. Thibaut, Depart-
ment of Psychology(Bat 32), 5 Boulevard du Rectorat, 4000Liège, Bel-
gium (e-mail: jpthibaut@ulg.ac.be).
647
Copyright 2002 Psychonomic Society, Inc.

648
THIBAUT, DUPONT, AND ANSELME
(67%). However, they did not replicateone result obtained as necessary features. In many cases, the fixed cate-
by Rips. In Rips, the similarity judgments were clearly in gories used by Rips (1989) or Smith and Sloman (1994)
favor of the fixed category (69% of choices), but not in were human activities defined by arbitrary rules or ob-
Smith and Sloman (50% in favor of the fixed category). jects defined by arbitrary values. In these situations, par-
In a second experimentalcondition(the rich condition), ticipants might have considered that even though the
Smith and Sloman (1994) added a characteristic feature defining feature is fixed by a rule (e.g., of the activity),
of the fixed category to the sparse description of the tar- in particular situations, it can be changed without modi-
get item (e.g., the characteristic feature “that is silver fying the nature of the activity (e.g., the official number
colored” was added to the original description “a circular of players—five—in a basketball game). These features
objectwith an X-inch diameter”). The purpose of this con- are not central in the sense defined by Sloman, Love, and
necessary-feature
dition was to contrast the
hypothesis Ahn (1998), for whom “feature centrality is a function of
characteristic-feature
with the
hypothesis. According to the extent to which other features depend on it ” (p. 191).
the necessary-feature hypothesis,participants categorize
In other experiments, dissociationswere obtainedin the
objectssolely on the basis of a necessary feature whenever context of a distinctionbetween deep and superficial fea-
one is available. The characteristic-feature hypothesis tures. The assumptionwas that categorizationis influenced
holds that people’s categorizations are based on charac- by “essential” (deep) features, whereas similarity is influ-
and
teristic
necessary features when both kinds are avail- enced by “superficial” (surface) features. In these studies,
able.In the rich condition,contrary to the sparse condition, participantswould use causal–explanatorytheories about
participants categorized the target items in the fixed cat- the world, and those theories would influence categoriza-
egory (i.e., the quarter, 77%) and judged them more sim- tion. People tend to hold essentialist beliefs about natural
ilar to this category(74%). Accordingto the authors, these kinds, and these categories are believed to have essential
results seem to corroborate the characteristic-feature hy- features that determine membership in the category. For
pothesis. Indeed, in the rich condition, participants cate- example, features such as “parents are dogs” or “gave birth
gorized the items more often in (and judged them as more to a zebra” or “has a certain genetic structure” are sup-
similar to) the fixed category than in the sparse condition. posed to be deep features, necessary for categorization.
This difference should not be expected if participants On the otherhand, superficial features (e.g., animal color)
were relying exclusivelyon the necessary feature. In sum, are supposed to be characteristic features that should in-
Smith and Sloman obtained a small dissociation only in fluence similarity judgmentsmore than categorization.In
the sparse condition and under think-aloud instructions. one experiment, Rips (1989) provided evidence of disso-
It is possible that many of the so-called characteristic ciation between categorization and similarity judgments
features used by Smith and Sloman (1994) in the rich con- by manipulating these two types of features. Participants
dition were not interpreted as more characteristic or less read stories about animals undergoing radical transforma-
essential than the hypothesized necessary features; com- tions. For example, a bird was transformed by a chemical
plementarily, it is also possiblethat the “necessary features” accidentinto somethingthat lookslike an insect.In thecat-
were not considered as more necessary than the character- egorization task, participants judged that the animal was
isticfeatures. Thismightbe duetotheway theauthorschose more likely to belong to the bird category, whereas in the
exemplars of both classes of features. The nature, “essen- similarity task they judged the animal to be more similar
tial” or “characteristic,” of the features was not established to an insect. Thus, such an experiment seems to corrobo-
independently but instead “by the authors’ judgment” rate the necessary-feature hypothesis, since participants’
(p. 379). It is possible that the size in the pizza-quarter ex- categorizationsfollowed deep features (“born from birds”)
ample was considered as highly implausible of the fixed rather than surface features. However, Estes and Hamp-
category of quarters. It is also possible that the added ton (2002) replicated Rips’s transformation study only
“characteristic” feature of the object was considered as a when they asked participants to perform both typicality
very implausiblepropertyof the stimuliof the variablecat- and categorizationjudgments(within design), whereas in
egory (e.g., being silvered for pizzas). Complementarily, a between-participants design there was no dissociation.
they could also have considered that “silvered” is as nec- Moreover, the authors found that even when the dissoci-
essary of quartersas the size of quarters is, in the sense that ation was obtained in Experiment 1, it was true of only a
a “nonsilvered”quarterwould be consideredas a false coin minorityof participants(in theirdata,less than30%). Thus,
(made of a wrong metal). After all, quarters have both again, thedissociationwas obtainedonlyin onecontextand
fixed size and material/color. Smith and Sloman notedthat only for a minority of people,which is not a very compel-
in the rich condition, when participantsmentioned a char- ling argument in favor of the necessary-feature hypothe-
acteristicfeature in their spontaneouscomments, they also sis (see also Hampton, 1995; Kalish, 1995; Malt, 1994).
categorized the stimulus in the fixed category. This result
However, these results should be interpreted cautiously.
is compatible with the authors’ characteristic-feature hy- First, these descriptionsrefer to very uncommon and im-
pothesisbut also with the idea that participantsdid not per- plausible events that, in turn, might lead to decreased
ceive the characteristic features as less necessary than the participants’ reliance on necessary features (e.g., a zebra
hypothesizednecessary features.
transformed into a new animal that behaves like a horse).
Consider now the possibility that the hypothesized Second, as mentioned by Hampton (1995), it is not obvi-
fixed-category-necessary features were not considered ous that participants really know and follow the assumed

SIMILARITY AND CATEGORIZATION
649
characteristic
difference between the necessary feature (genotype) and ing to the second category. Second,
features
the characteristic feature (phenotype).It is plausibleto as- were present in a subset of each category’s items. Third,
neutral
sume that most participants are not completely aware of we also introduced
features present in both cate-
the laws governingthe biologicalworld, and thus may ig- gories—that is, features that have no diagnostic role. The
nore which features are necessary (see also Ahn & Den- characteristic features were designed to be salient,
nis, 2001). If one cannot establish which properties a pri- whereas necessary features were not salient in the sense
ori categorizedas necessary (or characteristic)features by that a careful analysis was needed in order to find them.
the experimenterwere really interpretedand used as neces- Indeed, if the defining features were both salient and per-
sary (or characteristic) features by participants, one can- fectly predictiveand characteristicfeatures were notsalient
not decide which role is played by a property—necessary and not perfectly predictive, there would be no reason to
or characteristic—in a particular categorization.
rely on characteristic features either for categorization or
In sum, Rips (1989) found a dissociation, and Smith for similarity judgments. Another reason for contrasting
and Sloman (1994) replicated it in their sparse condition non-salient-defining features with salient-characteristic
under think-aloud instructions. Even in the latter case, features is that rule-based models and similarity-based
the dissociation was quite small and, as stressed by Estes models do not rely on these two classes of features in the
and Hampton (2002) in a similar context, it might have same way (see predictionsbelow).
been limited to a small percentage of participants.In each
We compared two experimentalconditionsthat differed
case, the authors hypothesized that participants were en- in terms of the association between the characteristic fea-
re-
gaged in some kind of analytic processing. The authors ture and the two categories. In the first condition (the
stricted
claimed that these results favor the characteristic-feature
condition in what follows), the characteristic fea-
hypothesis for categorization. However, we have sug- ture of Category A was absent in Category B and the
gested that the status of features, characteristic or neces- characteristic feature of Category B was absent in Cate-
sary, was questionable in these experiments.
The main purpose of our paper was to provideevidence it was a perfect cue, though limited, for one category.
cross-category
gory A. Each characteristicfeature was diagnosticbecause
that the dissociationsreally exist and to determinethe con-
In the second condition (hereafter the
text in which they appear. For that, we manipulated the condition),each characteristicfeature was associatedwith
role of necessary and characteristicfeatures. This is impor- one category in the same way as in the previous condition
tant since many categorizationmodels (such as prototype- except that it was also associated with one stimulus of the
and exemplar-based models) are similarity-based models other category. Thus, each characteristic feature was sta-
(Hampton, 1995, 1998; Nosofsky, 1986;Rosch, 1978)and tisticallydiagnosticthoughit was not a perfect cue for cat-
do not predict any dissociationbetween similarity and cat- egorization (see predictions below). This condition was
egorization.
introducedbecause with natural categories,and especially
Note that in Smith and Sloman’s (1994) sparse descrip- in experiments with artificial stimuli, when a (nondefin-
tion, where the dissociationwas obtained,stimuli were de- ing) feature is strongly associated with one category, it
scribed along only one dimension. The size in the pizza- also appears in stimuli of contrastive categories.
quarter example had the role of a necessary feature in the
The experiment was divided into a learning phase and a
fixed category and of a characteristic feature in the vari- test phase. In the test phase, new stimuliwere presentedfor
able category. This situationis not optimal for contrasting classification.The purpose of the experiment was to com-
the role of characteristic and necessary features in simi- pare similarity judgment and categorization obtained for
larity rating and categorization. One purpose of the pres- transfer stimuli, called “incongruent”—that is, built with
ent experiments was to provide evidence of dissociations the characteristicfeature of onecategoryand the necessary
with multidimensional stimuli by manipulating the fea- feature of the other category. Dissociationsshould occur if
tures constitutingstimuli. We created novelartificial stim- the two judgments do not rely on the same type of feature.
uli composed of features that appeared in an a priori fixed
Results obtained by Smith and Sloman (1994) and the
proportion of the stimuli: Stimuli were composed of fea- characteristic-feature hypothesis predict that participants
tures present in all the stimuli belonging to one category should not categorize solely on the basis of the necessary
(necessary features) and of features present in a subset of feature; some participantsshouldcategorizethe incongru-
stimuli belonging to one category (characteristic fea- ent stimuli according to the characteristic feature. More
tures). With stimuli perfectly controlled in terms of these generally, similarity-based approaches to categorization
constitutive features, one gets better control of the condi- also predict that characteristic features should influence
tions eliciting dissociations, if any.
categorization because they are salient (see models by
Hampton, 1995;Medin& Schaffer, 1978;Nosofsky, 1986;
Smith & Medin, 1981). According to these models, salient
characteristic features should have a large impact on simi-
EXPERIMENT 1
We designed stimuli made of two features, each fea- larity judgments and categorization. Similarity-based
necessary
ture being chosen among three types: first,
models predict that if there is a difference between the re-
features—that is, features that can be used as a rule for stricted and the cross-category conditions, categorization
categorization because they were present in each stimu- should be more influenced by the characteristic features
lus of one category and absent in all the stimuli belong- in the restricted condition.Thus, according to these mod-

650
THIBAUT, DUPONT, AND ANSELME
els, no dissociation should be obtained, especially in the cross-category condition, the cue validity of the characteristic fea-
ture “mushroom shape” was .87 for Category A and .13 for Cate-
restricted condition.By contrast, rule-based models pre-
gory B, whereas the reverse was true for the feature “angular.”
dict that categorization should be influenced only by the
In the test phase, 22 new stimuli, 11 per category, were con-
defining feature.
structed according to the same principles. For each category, there
were 2 congruent, 4 neutral, and 5 incongruent stimuli.
For similarity judgments on incongruent stimuli,
similarity-basedmodels predict that similarity judgments
should be driven by characteristic features, whereas rule-
based models have no strong prediction.
Procedure. Each participant was randomly assigned to one of
the two conditions—that is, 11 participants per condition. Partici-
pants were tested individually. They took 20– 45 min to complete
the task. The experiment was composed of two phases, a learning
phase and a test phase.
In the learning phase, participants were told that they would have
to learn to sort a set of stimuli into two categories. The initial stim-
ulus was presented to the participant, who had to guess its category
name (i.e., moffo and quipi). Feedback was provided about the ac-
curacy of the answers. The next stimulus was presented in the same
way, and so forth for the other stimuli. The order of presentation of
the stimuli was random. Once the entire set had been presented to
the participant, it was presented a second time, a third time, and so
on, until the participant made no mistake during two successive pre-
sentations of the set of the stimuli.
Dissociations should occur if participants use the non-
salient necessary feature for categorization and the char-
acteristicsalientfeature for similarityjudgment(or the re-
verse, which is unlikely). According to similarity-based
models and Smith and Sloman’s (1994) hypothesis, we
shouldnot expect such dissociationsto occur. By contrast,
rule-based models have no strong prediction about disso-
ciations because they do not specify the role of character-
istic features in similarity judgments. They would predict
dissociationsonly if one assumes that even though people
use a rule for categorization,they also registerselected as-
In the test phase, participants were presented with the test stim-
sociations between a feature and a category, especially uli one at a time. For each stimulus, the participant was asked to de-
cide which of the two categories the stimulus belonged to and to
when the associated feature is salient.
choose the category the object was more similar to. Half of the par-
Anotherpurposeof the experimentwas to show thatdis-
ticipants performed the similarity task first, whereas the other half
sociations are not the result of the use of deep causal fea-
performed the categorization task first. This first task was followed
by a rating task. Participants had to rate on a scale from 1 to 7
tures (e.g., a genetic cause) for categorizationand superfi-
cial features for similarity judgments. If dissociations
come through deep features, no dissociationshould be ob-
tained here given that our features had no deep causes but
only statistical regularities.
whether the test stimuli were likely to belong to Category A or to
Category B. Similarly, they also had to rate whether the test stimuli
were more similar to Category A or to Category B. The end of the
scale corresponding to 1 referred to Category A and the end corre-
sponding to 7 referred to Category B.
Method
Results
Participants. Twenty-two undergraduate students from the Uni-
versity of Liège volunteered for the experiment.
First, we searched for dissociationsbetween categoriza-
tion and similarity judgments. As predicted, there was no
dissociationfor congruentstimuli. We analyzedthe results
Materials. In the learning phase, two categories of 10 artificial
stimuli were constructed. The stimuli were novel shapes that were
composed of two parts, an upper part (the different F2 parts in Fig-
ure 1) and a lower part (the F1 parts in Figure 1). As mentioned above,
there were two conditions, the restricted and the cross-category con-
ditions. In the restricted condition, for 6 stimuli out of 10, the upper
part, for Category A, has a mushroom shape slightly distorted over
the 6 stimuli (F2a), and an angular shape for stimuli in Category B
1
obtained for the 10 incongruent test stimuli (i.e., 5 F1a
1
F2b stimuli and 5 F1b F2a stimuli). We considered that
a participant dissociated categorization and similarity
judgment when he/she categorized 9 or 10 test stimuli in
one category while estimating them more similar to the
(F2b). These 6 stimuli were called “congruent” (panel A displays other category. Twelve participants (out of 22) produced
the 6 congruent stimuli of each category). The 4 remaining stimuli
such a dissociation. However, a comparison between the
of the two categories, called the “neutral” stimuli, were constructed
restricted and the cross-category conditionsrevealed that
with four different upper parts (F2c,d,e,f ). F2c,d,e,f are present in
10 participants (out of 11) dissociated in the restricted
both categories and thus they cannot be used as cues for catego-
condition and only 2 in the cross-category condition. A
rization (panel B displays the 4 neutral stimuli from the two cate-
Fisher exact test revealed that the proportion of dissoci-
ationsobtaineddiffered significantlyin the two conditions
gories). Each lower part is composed of four legs that are spatially
grouped as one leg on the left and three legs on the right for Cate-
gory A (1–3 legs; see Feature F1a in the stimuli), and two sets of
two legs in Category B (2–2 legs; see F1b). The cross-category con-
dition was constructed in the same way except that 1 of the 4 “neu-
tral” stimuli from Category A and 1 neutral stimulus from Cate-
gory B (SA10 and SB10, respectively) were replaced by 2 new
stimuli. The first new stimulus was composed of the necessary
(F1a) feature of Category A (i.e., 1–3 legs) and the characteristic
p ,
(
.01). To summarize, this analysis revealed that there
were more dissociationsbetween categorizationand sim-
ilarity judgments in the restricted condition than in the
other condition.
In order to confirm these analyses, a two-way analysis
3
of variance (ANOVA) (2 2) with category type (cross-
¢
feature of Category B (F2b); this stimulus was called “SA10 .” The
category and restricted) as a between-participants vari-
able and task (categorization and similarity) as a within-
participantsvariablewas performed ontheratingsobtained
for similarity judgments and for categorization. Dissoci-
ations are obtained for stimuli that get a small score for
categorization and a high score for similarity or a small
second stimulus was composed of one necessary feature from Cat-
egory B (F1b) (2–2 legs) and one characteristic feature of Cate-
¢
gory A (F2a), a stimulus called “SB10 ” (Figure 1C). We call these
2 contradictory stimuli “incongruent.”
In sum, the cue validity of each characteristic feature (e.g., the
probability that a stimulus is a member of Category A given that it
has a mushroom shape) was 1 in the restricted condition. In the score for similarity and a high score for categorization.

SIMILARITY AND CATEGORIZATION
651
Figure 1. Panels A and B: The 20 stimuli from the restricted condition. Panel A: The 12 stimuli from Category A (SA1–SA6) and
Category B (SB1–SB6) used in the learning phase. Panel B: The 8 neutral stimuli, SA7–SA10 for Category A, and SB7–SB10 for Cat-
¢
¢
egory B. Panel C: In the cross-category condition, SA10 and SB10 were replaced by the “contradictory” stimuli SA10 and SB10 .
In order to perform a single analysis on the scores ob- a participantcategorizeda stimulusin—or estimated it as
tained for the test stimuli from both categories (1–3 and more similar to—the category defined by the necessary
2–2 test stimuli scores), we recoded the categorization feature, whereas a high score indicates that the stimulus
and similarity judgment scores obtained for the 2–2 in- has been categorized in—or estimated more similar to—
congruent test stimuli (i.e., stimuli that had to be cate- the category defined by the characteristic feature. Results
gorized in B when the participantfollowed the 2–2 rule). are shownin Figure 2. There was a significanteffect of con-
F
5
p ,
High scores in categorization,indicatingthat participants dition [ (1,20) 33.96, .0001; results obtained in
categorized the 2–2 stimuli in Category B, were trans- the restricted conditionwere higher than the equivalentre-
formed into small scores and vice versa. Small scores for sult in the cross-category condition,3.45 vs. 1.58, respec-
F
5
X 5
p ,
X 5
3.9 for
similarity judgment, indicating that participants judged tively], of task [ (1,20) 63.51,
.0001;
1.14 for categorization],
p ,
the 2–2 test stimuli as more similar to Category A, were similarity judgments vs.
transformed into high scores and vice versa. In other and a significant interaction [ (1,20)
words, after transformation, a small score indicates that .0001; see Figure 2]. A posteriori test revealed a signifi-
F
5
25.77,

652
THIBAUT, DUPONT, AND ANSELME
cant difference between the restricted and cross-category tions, characteristic features did not influence catego-
p
conditionsfor similarity judgments (Tukey HSD, < .05; rization.
2.08 vs. 5.7 in the cross-category and the restricted con-
In the cross-category condition, even similarity judg-
ditions, respectively). There was no difference between ments were not influencedby characteristic features. This
these two conditionsfor categorization(1.08 vs. 1.2 for the suggeststhat participants’similarityestimationwas driven
cross-category and the restricted conditions,respectively). by defining features only, by contrast with what was pre-
We computed a confidence interval on the similarity and dicted by the above theories. However, one should inter-
the categorizationscores in the two conditionsat the level pret the cross-category results cautiously. One possibility
a
of = .05.The confidenceintervalsfor the categorization is that, because of the incongruentstimuli, participantsdid
scores were [0.81, 1.35] in the cross-category condition not notice the associationbetween each characteristicfea-
and [0.93, 1.47] in the restricted condition. For the simi- ture and one of the two categories. Given that the charac-
larity ratings, they were [1.06, 3.10] in the cross-category teristic feature was salient, it was, most likely,the first fea-
condition and [4.69, 6.72] in the restricted condition. In ture tested by participants. Suppose that this first rule
the restricted condition, the hypothesis that the mean worked for a couple of trials; then, one contradictoryitem
would be beyond the value 4 (the intermediate value be- appeared, refocusing the participant’s attention to other
tween 1 and 7 on the scale) was rejected for the catego- potentialrelevant features. For the rest of the experiment,
rization scores whereas it was accepted for the similarity even if participants noticed the presence of each charac-
scores.
teristic feature, they would not pay attention to the cate-
In order to specify the lack of influence of characteris- gory with which it was associated. The interspersed neu-
tic features on categorization, we compared the catego- tral features (6 stimuli) also obscured the relationship
rization ratings obtainedfor the congruentand the neutral between each characteristic feature and its category. In
stimuli in both conditions.If the characteristic feature in- sum, if this were true, the absence of influence of charac-
fluenced categorization,we should get smaller ratings for teristic features on the similarity judgments in the cross-
congruent than for neutral stimuli. In both conditions, category condition would not be the result of a rigid rule-
there was no significant difference between the two types based behavior (driven by the presence of contradictory
of stimuli.
items) but the result of a misperception of the association
between onecharacteristicfeatureand a particularcategory.
DISCUSSION
EXPERIMENT 2
The analysis of the confidence interval confirms the
results obtainedfor the ratings. The dissociationbetween
similarity and categorization appeared, for the majority
of participants,in the restricted condition,whereas no dis-
sociation was obtained in the cross-category condition.
The two conditions differed in terms of the similarity
scores but not in terms of the categorization scores. In
both conditions, categorization was based on the rule,
whereas similarity judgments were based on the charac-
teristic features in the restricted conditionand on the rule
in the cross-category condition. Our results also show
that it is possible to obtain dissociations even when there
is no deep, causal feature (vs. surface features) involved
in the experimental design. The data did not confirm the
characteristic-feature hypothesis since, in both condi-
In Experiment 2, a new conditionwas created in which,
before learning, participants were asked to free sort the
set of stimuli into two categories. A free-sorting task
should give them an overview of the distribution of each
characteristic feature in the stimuli and, because of their
salience,participantsshould use them to free sort the stim-
uli, and later use them as features relevant for categoriza-
tion at the beginning of the learning phase. If in the first
experiment participants did not dissociate in the cross-
category condition because they did not notice the asso-
ciation, they should dissociate in the present experiment,
as in the restricted condition. On the other hand, if they
had noticed the existence of the association in the cross-
category condition of the first experiment and used nec-
essary features for categorization and similarity judg-
ment because of the presence of the contradictory items
in the learning phase, again, we should obtain no disso-
ciation between similarity and categorization. The re-
stricted condition was also run in the same way as a con-
trol condition.
Method
Participants. Twenty-two undergraduate students from the Uni-
versity of Liège participated as volunteers in the experiment.
Material. The stimuli used in the first experiment were also used
here. The two conditions compared in the present experiment were
defined in the same way as in the first experiment: In the restricted
condition, 6 congruent stimuli and 4 neutral stimuli and, in the cross-
category condition, 6 congruent stimuli, 3 neutral stimuli, and 1 in-
Figure 2. Interaction task 3 condition. Note the dissociation
between categorization and similarity judgmentsin the restricted
condition.

SIMILARITY AND CATEGORIZATION
653
congruent stimulus. The 22 test stimuli, 11 per category, used in the That is, after transformation, a low score indicates that a
first experiment, were also used in the present experiment.
Procedure. Each participant was randomly assigned to one of the
participant categorized a stimulus in—or judged it as
more similar to—the category defined by the necessary
two conditions, for 11 participants per condition. Again, participants
feature, whereas a high score indicates that the stimulus
were tested individually. The experiment was composed of three
has been categorized in—or judged more similar to—the
phases: a free-sorting task, a learning phase, and a test phase. In the
category defined by the characteristic feature. There was
free-sorting task, participants were presented with the 20 stimuli de-
signed for the learning phase and asked to sort the stimuli into two
categories, the only constraint being that each category should com-
prise 10 stimuli. Then, participants moved to the learning phase and
the test phase, which were the same as in Experiment 1.
F
5
p ,
a significant effect of task [ (1,20) 108.28,
.0001;
5.37 for similarity
judgments]. There was no significant effect of category
p .
X 5 X 5
1.34 for categorization vs.
F
5
type [ (1,20) 0.0,
.1], and, by contrast with Experi-
p .
F
5
ment 1, no significantinteraction[ (1,20) 0.01,
.1].
Results and Discussion
As in Experiment 1, we computed a confidence interval
Free-sorting task. Two participants found the correct
rule for categorization (1–3 vs. 2–2). Eighteen partici-
pants used the upper part as a distinctive cue (mushroom
vs. angular shape). They categorized the two members of
a neutral pair (e.g., SA7 and SB7) in the same category,
and the members of another neutral pair in the second
category. Since in the cross-category conditionthere were
onlythree neutral pairs, participantsput 1 stimulusin each
category in order to have the same number of stimuli in
each category as required by the experimenter. The last 2
participantsmentionedthat they categorizedaccordingto
the shape of the legs: their length,their thickness,or both.
Test phase. The main purpose of the present experi-
ment was to replicate the results obtained in the restricted
condition in the first experiment and to generalize these
results to the cross-category condition when participants
first analyzed stimuli in the free-sorting task. The number
of dissociations obtained in the two conditions for the 10
on the similarity and the categorization scores in the two
a
5
.05. The confidence inter-
conditions at the level of
vals were [1.05, 1.60] for the categorization scores in the
cross-category conditionand [1.08, 1.63] in the restricted
condition.For the similarity ratings, they were [4.33, 6.36]
in the cross-category condition and [4.36, 6.40] in the re-
stricted condition. These results show a dissociation in
both conditions. In fact, the hypothesis that the mean is
beyondvalue 4 (the intermediatevalue between 1 and 7 on
the scale) was rejected for the categorization scores
whereas it was confirmed for the similarity scores.
As in Experiment 1, we compared the categorization
ratings obtained for the congruentand the neutral stimuli
in both conditions in order to assess the role of charac-
teristic features on categorization. In both conditions,
there was no significantdifference between the two types
of stimuli. For similarity ratings, note that the difference
between congruent and neutral stimuli was significant in
1
incongruent test stimuli (i.e., 5 F1a F2b stimuli and 5
F
5
p ,
F
5
both conditions[ (1,10) 73.8,
.001, and (1,10)
.001 for the cross-category condition and the
1
F1b F2a stimuli) was compared. A participant was cat-
p ,
53.83,
egorized as a dissociator when he/she categorized 9 or 10
incongruenttest stimuli in one of the two categories (A or
B) while judging them more similar to the other category
(B or A). Nine participants (out of 11) produced such a
dissociation in the cross-category condition. The same
number of participants were dissociators in the restricted
condition.The proportionof dissociationsobtainedin the
cross-category condition was also compared to the corre-
sponding proportion (i.e., 2 dissociators) obtained in this
condition in the first experiment. A Fisher exact test re-
restricted condition, respectively], where neutral stimuli
got higher scores than congruent stimuli, indicating that
they were judged as less similar to their category mem-
bers than were congruent stimuli. This is consistent with
the idea that characteristic features influenced similarity
judgments.
The present experiment replicates the results obtained
in the restricted condition in the first experiment. The ex-
periment also revealeda dissociationin the cross-category
condition, contrary to what happened in Experiment 1.
This suggests that the absence of dissociation in the first
experimentwas due to participantsfailingto noticethe as-
sociation between each characteristic feature and a cate-
gory. Note that the result is potentiallysurprising because
free-sorting tasks are notorious for leading people to one-
dimensional sortings (Medin, Watenmaker, & Hampson,
1987; Spalding & Murphy, 1996). Thus, the prediction
that participants would notice that the dimension they
used in the free-sorting task was associated with other di-
mensions was not obvious because, after all, if they no-
ticed the association, they should use it in their sorting.
p ,
vealed that the two proportionsdiffered significantly(
.01). Thus, the results revealed that, with a free-sorting
task added before the learning task, most of the partici-
pants dissociated categorizationand similarity judgments
in the cross-category condition.
3
As in the first experiment, a two-way ANOVA (2 2)
with category type (cross-category and restricted) as a
between-participantsvariable and task (categorizationand
similarity) as a within-participantsvariablewas performed
on the ratings obtained for similarity judgments and for
categorization. Dissociations were obtained for stimuli
with a small score for categorization and a high score for
similarity or a small score for similarity and a high score
for categorization. As in the first experiment, we recoded
the categorization and similarity scores obtained for the
GENERAL DISCUSSION
Our main purpose was to study the role of characteristic
2–2 incongruenttest stimuli(i.e., stimulithat had to be cat- and defining features in similarity judgments and catego-
egorizedin B when the participantfollowed the 2–2 rule). rization, particularly in the case of dissociations between

654
THIBAUT, DUPONT, AND ANSELME
two tasks(Rips, 1989;Smith & Sloman, 1994).Our results in both similarity and categorization judgments and that
showed that, when the frequency of association between cause features matter more than effect features in cate-
characteristic and necessary features and each category gorization. They concluded that there is a tendency to
was controlled, no participant categorized incongruent give more weight to surface features than people would
stimuli in the category associated with the characteristic in categorization judgment and, in that sense, catego-
feature, therebysuggestingthat participantsdid not use the rization does not depend on similarity (see also Medin &
characteristic feature in their categorizations.By contrast, Ortony, 1989; Rips, 1989). Our results show that this
similarity judgments were strongly influenced by charac- statement is restrictive in that dissociations can be ob-
teristic features. We will discuss these results in the lightof tained in the absence of deep features.
the characteristic-feature hypothesis and similarity-based
models and in the context of the aforementioned associa- Why Participants Dissociate Similarity
tion between dissociations and deep features.
Judgments and Categorization
In their experiments, Smith and Sloman (1994) ob-
Estes and Hampton (2002) suggested that dissociations
served a small dissociationin their sparse conditionunder were obtained in situationsthat favor some kind of “reflec-
think-aloud instructions and no dissociation in their rich tive processing.” This account is also compatible with
condition.By contrast, we obtaineda dissociationin a sit- Smith and Sloman’s (1994) result that dissociations were
uation that was at least as rich as their rich condition, obtainedonly when participantswere engaged in a think-
showing that dissociations can be obtained with multidi- aloudsituation.As suggestedby Estes and Hampton, these
mensional stimuli and for a majority of participants. As situationsmay promote a kindof contrastiveprocessing,in
mentioned in the introduction, one possible reason why the sense that participants who are asked to make two (or
the authors did not get a sharper dissociationwas that par- three) successive judgments may seek ways to differenti-
ticipants did not interpret the features as characteristic or ate them and thus be induced to dissociate their ratings.
necessary, contrary to what they were supposed to be. This does not explain why people always dissociate in the
Note also that our data do not confirm the characteristic- same direction—that is, why similarityjudgmentsare more
feature hypothesis proposed by the authors.
influenced by surface features and categorization is more
Second, a number of authors have suggested that the influenced by deep features. Also, the contrastive hy-
dissociations should be obtained when surface and deep pothesis does not explain the dissociation obtained by
features define the experimental situation (Kroska & Ahn and Dennis (2001) using a between-participants de-
Goldstone,1996;Rips, 1989;see, however,Estes & Hamp- sign. Our designwas a within-participantsdesignand thus
ton, 2002; Hampton, 1998). The introductionstressed that our dissociationscould be the result of the contrasting hy-
the status of the features, either deep or surface, was often pothesis. However, recall that half of the participantswere
debatable and this made the interpretation of the results asked to perform the similarity judgments first, whereas
less straightforward. Recently, Ahn and Dennis (2001) the other half had to begin with categorization. Consider
conducted experiments in which the status of the fea- participants who dissociated. If we compare the catego-
tures— deep versus superficial features (in the authors’ rization results of those who started with categorization
terms, cause features vs. effect features)—was controlled with the similarity judgment results of those who started
and dissociations were obtained. According to the causal with similarity judgments, we have a perfect between-
status hypothesis (Ahn, 1998), features that cause other participants dissociation because all the categorization-
features receive more weight in categorization judgment first participantscategorizedaccordingto the definingfea-
andintheframework of psychologicalessentialism(Medin ture and all the similarityjudgment-firstparticipantsused
& Ortony, 1989). The authors hypothesized that catego- the characteristic feature in their estimation of similarity.
rization would be more influenced by deep features than
We believe that dissociations will appear each time the
similarity judgments. In their paradigm, a target situation contrasted categories have features that people interpret
is defined by three features: A that causes B that causes C. as defining (even if these features are not really defining),
Consider now two situations, one sharing a single deep whereas they interpret other features—the characteristic
feature (Feature A) with the target, the other situation features—as associatedwith one particularcategorywith-
sharing two surface features with the target (Features B out being defining of this category. Because of the learn-
and C). If the deeper features are weighted more heavily ing phase, this was the case with our characteristic and
in categorizationjudgments than in similarity judgments, defining features. In terms of the control of features, the
the first situation should be selected more frequently in paradigm used by Ahn and Dennis (2001) is related to
the categorization task than in the similarity judgment ours. They achieved the difference between deep and sur-
task. The results confirmed this view. Though deeper fea- face features through the manipulation of the causal con-
tures influenced similarity judgments (56% of the partic- nections between features, whereas in our situation the
ipants judged the situation sharing one deep feature with defining and characteristic features got their status during
the target more similar to the target), categorization was the learning phase.
more influenced by these deep features (75% of the par-
In order to explainthe orientationof dissociations(sim-
ticipants categorized the one-deep-feature-shared situa- ilarity goes with surface or characteristicfeatures and cat-
tion with the target). Ahn and Dennis argued that feature egorization with deep or necessary features), one has to
weighting is determined by the causal status of features assume that people weight features differently depend-

SIMILARITY AND CATEGORIZATION
655
ing on the task. Why would this be the case? According why this may have been the case. We would say that un-
to Ahn and Dennis (2001), a mechanism underlying their less participantsare pushed to look beyond the immediate
dissociation is that when deep features are defined as situation, they have no reason to search for less obvious
those that cause surface features, deep features are features. By contrast, instructionssuch as “considerall the
weighted more in categorization judgments than in sim- information available” may motivate them to search for
ilarity judgments. The reason for this difference in features that are not directly available in the stimuli (see
weighting is, according to the authors, that people be- Medin, Goldstone, & Gentner, 1993; Thibaut & Schyns,
lieve that the surface features of the members of a cate- 1995, for discussion of the notion of setting a feature
gory share the same underlying cause. On the other space for categorization and similarity judgments).
hand, unless the goal of the similarity judgment is spec-
ified, the relative weighting one feature gets depends on Dissociations and Similarity Models
and varies with context and task. In our case, catego- of Categorization
rization was not associated with deep features. Thus, we
As mentioned in the introduction,many models of cat-
have to rephrase this statement in more general terms: egorization rely on the idea that categorizationis a matter
Features that are thought to be perfectly predictive of a of similaritycomputation.This idea is central in prototype
category with respect to the other possible categories models that assume that categorizationis a matter of com-
will receive more weight in categorization than other parison between the target and prototypesstored in mem-
features. Features used in similarity judgments depend ory. The idea is also central for exemplar-based models
more on the context. Since the number of possible con- (Kruschke, 1992; Nosofsky, 1986). These similarity-
texts is infinite (Murphy & Medin, 1985), there is a large based models do not account for our results. For example,
number of features that can be referred to in similarity consider Nosofsky’s (1986) generalized context model or
judgments.
Kruschke’s (1992) ALCOVE model. According to these
The importance of defining features in the case of cat- models, participants learning the categories in our exper-
egorization is illustrated by our restricted condition. In iments will shift their attention to the leg dimension be-
this condition,both types of features were perfect cues for cause it perfectly predicts the categories. This feature will
categorization, the only difference being that the defin- be heavily weighted in the classification decision—and
ing feature was associated with all the stimuli in one cat- because the classification decision is based on similarity,
egory, the characteristicfeature beingassociatedonly with the similarity between an instanceand a categorywill also
a subset of one category stimuli. In this case, categoriza- be strongly affected by this dimension, which is in con-
tion was not influencedby characteristicfeatures (see also flict with our data. On the other hand, if participants start
the absence of difference between the ratings obtainedfor first with the salientcharacteristicfeature, in the restricted
the congruent and the neutral stimuli).
condition the characteristic feature always points to the
One also has to explain why participantsdid not choose same category; even if the defining feature gets more at-
the same features in both tasks. We believe that when tention as learning proceeds, there is no reason why this
there is no particulargoal specified for the similarity judg- feature would not affect categorization,whereas it remains
ment, participants will use features that encompass the central for similarity judgments. Clearly, characteristic
largest part of the compared entities. In our situation, the features did not affect categorization since there was no
characteristic feature was the most salient one because it difference between neutral and congruent items.
constitutedvirtually the totality of the stimulus. In a sim-
Recently, a number of categorization models have at-
ilarity judgment for which features must be chosen, it tempted to classify instances using rules and similarity
may seem natural to rely on features that are the most to previously experienced exemplars (see Erickson &
salient and to neglect other features, unless there is some- Kruschke, 1998; Nosofsky, Palmeri, & McKinley, 1994).
thing in the task that requires going beyond the most ob- Hybrid models should explain why participants learned
vious features available.Thus, if the defining feature had nothing about the association between the characteristic
been more salient than the characteristic feature, no dis- feature and its category in the cross-category condition,
sociation would have been obtained (in Experiment 2, or why the contradictorystimuliencounteredduringlearn-
most participants in the free-sorting task spontaneously ing did not influence the test phase. Indeed, these items
sorted stimuli according to the salient characteristic fea- presented during learning should be weighted as specific
ture, whereas a small minority used the defining feature). exemplars belonging to one category and influence cat-
Similarly, in Ahn and Dennis (2001), participantsweighted egorizationof the contradictory items at test. One should
the two characteristic features more in similarity judg- obtain a difference between the cross-category condition
ments than in categorization. Interestingly, they showed and the restricted condition for the contradictory items,
that when the similarity instructionswere “Considerall of which was not confirmed by the data.
the information available. Which option (A or B) is the
In general, in similarity models, the weight of each fea-
target more like?” instead of the standard instructions, ture is a matter of its diagnosticityand saliencein the cat-
there was no dissociation between similarity and catego- egory with which it is associated (Thibaut & Schyns,
rization. Relatedly, Goldstone (1994) argued that the task 1995). To capture the dissociations studied here, one
demand may force participants to interpret “similar” as should add a mechanism that emphasizes the role of fea-
“visually similar.” However, he did not clearly explain tures that are interpreted as necessary by people (even

656
THIBAUT, DUPONT, AND ANSELME
though these features are not very salient) and that de- Medin, D. L., & Schaffer, M. (1978). Context theory of classification
learning. Psychological Review, 85, 207-238.
Medin, D. L., Watenmaker, W. D., & Hampson, S. E. (1987). Family
resemblance, conceptual cohesiveness, and category construction.
Cognitive Psychology, 19, 242-279.
Murphy, G. L. (2002).The big bookof concepts. Cambridge, MA: MIT
Press.
crease the importance of salient characteristic features in
the case of categorization, whereas similarity judgments
seem to be determined more by the salience of the fea-
tures involved in the comparison.
Murphy, G. L., & Medin, D. (1985). The role of theories in conceptual
coherence. Psychological Review, 92, 289-316.
REFERENCES
Nosofsky, R. M. (1986). Attention, similarity, and the identification-
categorization relationship. Journal of Experimental Psychology:
General, 115, 39-57.
Ahn, W. K. (1998). Why are different features central for natural kinds
and artifacts? The role of causal status in determining feature cen-
trality. Cognition, 69, 135-178.
Nosofsky, R. M., Palmeri, T. J., & McKinley, S. C. (1994). Rule-plus-
exception model of classification learning. Psychological Review,
101, 53-79.
Rips, L. J. (1989). Similarity, typicality, and categorization. In S. Vos-
niadou & A. Ortony (Eds.), Similarity and analogical reasoning
(pp. 21-59). Cambridge: Cambridge University Press.
Rips, L. J., & Collins, A. (1993). Categories and resemblance. Journal
of Experimental Psychology, 122, 468-486.
Ahn, W. K., & Dennis, M. J. (2001) Dissociation between categoriza-
tion and similarity judgement: Differential effect of causal status on
feature weights. In U. Hahn, & M. Ramscar, (Eds), Similarity and
categorization (pp. 87-107). New York: Oxford University Press.
Erickson, M. A., & Kruschke, J. K. (1998). Rules and exemplars in
category learning. Journal of Experimental Psychology: General,
127, 107-140.
Est es, Z., & Hampton, J. A. (2002). Similarity and essentialism in cate-
gory judgmentsof naturalkinds. Manuscript submitted for publication.
Goldstone, R. (1994). The role of similarity in categorization: Pro-
viding a groundwork.Cognition, 52, 125-157.
Rosch, E. (1978). Principles of categorization. In E. Rosch & B. B.
Lloyd (Eds.), Cognition and categorization (pp. 27-48). Hillsdale,
NJ: Erlbaum.
Sloman,S. A., Love, B. C., & Ahn, W. K. (1998). Feature centrality and
conceptual coherence. Cognitive Science, 22, 189-228.
Smith, E. E., & Medin, D. L. (1981). Categories and concepts. Cam-
bridge, MA: Harvard University Press.
Hampton, J. A. (1995). Testing the prototypetheory of concepts. Jour-
nal of Memory & Language, 34, 686-708.
Hampton, J. A. (1998). Similarity-based categorization and fuzziness
of natural categories. Cognition, 65, 137-165.
Smith, E. E., & Sloman, S. A. (1994). Similarity- versus rule-based
categorization. Memory & Cognition, 22, 377-386.
Spalding,T. L., & Murphy, G. L. (1996).Effects of backgroundknowl-
edge on category construction. Journal of Experimental Psychology:
Learning, Memory, & Cognition, 22, 525-538.
Thibaut, J. P. (1997).Similaritéet catégorisation. Année Psychologique,
97, 701-736.
Thibaut, J. P., & Schyns, P. G. (1995). The development of feature
spaces for similarity and categorization. Psychologica Belgica, 35,
167-185.
Kalish, C. W. (1995). Essentialism and graded membership in animal
and artifact categories. Memory & Cognition, 23, 335-353.
Komatsu, L. K. (1992). Recent views of conceptual structure. Psycho-
logical Bulletin, 112, 500-526.
Kroska, A., & Goldstone, R. L. (1996). Dissociations in the similarity
and the categorization of emotions. Cognition & Emotion, 10, 27-46.
Kruschke, J. K. (1992). ALCOVE: An exemplar-based connectionist
model of category learning. Psychological Review, 99, 22-44.
Malt, B. C. (1994). Water is not H2O. Cognitive Psychology, 27, 41-70.
Medin, D. L., Goldstone, R. L., & Gentner, D. (1993). Respects for
similarity. Psychological Review, 100, 254-278.
Medin, D. L., & Ortony, A. (1989). Psychological essentialism. In
S. Vosniadou & A. Ortony (Eds.), Similarity and analogicalreason-
ing (pp. 179-195). Cambridge: Cambridge University Press.
(Manuscript received September 27, 2000;
revision accepted for publication January 9, 2002.)