ANNUAL

REVIEWS Further
Quick links to online content

Ann. Rev. Psychol. 1981. 32:89-115

Copyright © 1981 by Annual Reviews Inc. All rights reserved

CATEGORIZATION OF +341

NATURAL OBJECTS
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

Carolyn B. Mervis

Department of Psychology, University of Illinois, Champaign, Illinois 61820

by University of Sussex on 10/12/12. For personal use only.

Eleanor Rosch

Department of Psychology, University of California, Berkeley, California 94720

CONTENTS

THE NONARBITRARY NATURE OF CATEGORIES .......................................... 91

Basic Level Categories ..... . . . ........ ........ .. .... ........................... ................................
. . . .. . . 92
Basic Level Categorization as a Basic Process .. ............. ... . . .. . ........ . . .
. . .. ... . . . ......
. . . ..... 94
NONEQUIVALENCE OF CATEGORY MEMBERS................................................ 95
INDETERMINACY OF CATEGORY MEMBERSHIP
AND REPRESENTATION .......................................................................................... 100
THE NATURE OF ABSTRACTION .......................................................................... 102
DECOMPOSABILITY OF CATEGORIES INTO ELEMENTS .............................. 104
THE NATURE OF ATTRIBUTES.............................................................................. 106
SUMMARY ........................................................... ......................................................... 109

A category exists whenever two or more distinguishable objects or events

are treated equivalently. This equivalent treatment may take any number
of forms, such as labeling distinct objects or events with the same name, or
performing the same action on different objects. Stimulus situations are
unique, but organisms do not treat them uniquely; they respond on the basis
of past learning and categorization. In this sense, categorization may be
considered one of the most basic functions of living creatures.
The last chapters on concept formation in the Annual Review of Psy­
chology (Neimark & Santa 1975. Erickson & Jones 1978) treated concepts
(categories) as part of the study of problem solving within the general field
of psychological learning theory. Meanwhile, an essentially new field of
research and theory concerning concepts and categories has emerged, fed

89
0066-4308/8110201-0089$01.00
 90 MERVIS & ROSCH

by two major trends: 1 . The study of naturalistic categories (for example,

"red," "chair") particularly as influenced by input from anthropology,
philosophy, and developmental psychology. 2. The modeling of natural
concepts in the field called semantic memory, an area greatly influenced by
artificial intelligence. This chapter is a selective review of these newer
developments.
First a word about the historical context of this work. In the usual way
of thinking, people distinguish objects (material things in space and time)
from the attributes, properties, or qualities of those objects (such as color,
shape, function, or parts). In the history of thought, there have been many
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

ideas about the nature of objects, of qualities, of their relation to one

another, and of their relation to the ideas which people have of them. British
empiricist philosophy had one such view. In British empiricism (for exam­
by University of Sussex on 10/12/12. For personal use only.

ple, Locke 1 690), the ideas that people have of objects (concepts) consist
of an intension (meaning) and an extension (the objects in the class). The
intension is a specification of those qualities that a thing must have to be
a member of the class; the extension consists of things that have those
qualities. Thus, qualities (attributes) connect concepts to the world. Con­
cept formation research within learning theory, by the very nature of its
research paradigm, presupposes a British empiricist stand on these issues
(e.g. see Fodor 198 1).
In a classical concept formation experiment, stimuli are typically sets
of items varying orthogonally on a limited number of sensory qualities
such as color and form. Concepts are complexes composed of and decom­
posable into the defining qualities and logical relations between those qual­
ities (e.g. red and square) which are their elements. Originally, the passive
and gradual learning of common defining elements was emphasized in re­
search (Hull 1920). However, since Bruner (Bruner et al 1956), research
has concentrated on subjects' active hypothesis testing in the learning
of relevant features and the logical rules combining them (see Bourne et
al 1979).
The newer categorization research has raised for debate at least six
empirical and theoretical issues, none of which had been considered debat­
able by the earlier approach. They are listed below, beginning with the
somewhat more concrete structural problems.

1. Arbitrariness of categories. Are there any a priori reasons for dividing

objects into categories, or is this division initially arbitrary?
2. Equivalence of category members. Are all category members equally
representative of the category, as has often been assumed?
3. Determinacy of category membership and representation. Are categories
specified by necessary and sufficient conditions for membership? Are
boundaries of categories well defined?
 CATEGORIZATION OF OBJECTS 91

4 . The nature o f abstraction. How much abstraction i s required-that is,

do we need only memory for individual exemplars to account for catego­
rization? Or, at the other extreme, are higher-order abstractions of gen­
eral knowledge, beyond the individual categories, necessary?
5. Decomposability of categories into elements. Does a reasonable explana­
tion of objects consist in their decomposition into elementary qualities?
6. The nature of attributes. What are the characteristics of these "at­
tributes" into which categories are to be decomposed?

Below we consider work bearing on each of these issues.

Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

THE NONARBITRARY NATURE OF CATEGORIES

by University of Sussex on 10/12/12. For personal use only.

In the stimulus sets of the classical concept formation paradigm, attributes

are combined arbitrarily to form items. This view has been echoed in related
disciplines " . . . the physical and social environment of a young child is
perceived as a continuum. It does not contain any intrinsically separate
'things' " (Leach 1964, p. 34).
However, the contention that the division of real world objects into
categories is originally arbitrary would make sense only if the attributes in
the world formed a total set (in Garner's 1 974 sense); that is, if all combina­
tions of attribute values were equally likely to occur. For example, consider
some of the qualities ordinarily treated as attributes in classifying animals:
"coat" (fur, feathers), "oral opening" (mouth, beak), and "primary mode
of locomotion" (flying, on foot). If animals were created according to the
total set model, then there would be eight different types:

(a) those with fur and mouths, which move about primarily on foot;
(b) those with fur and mouths, which move about primarily by flying;
(c) those with fur and beaks, which move about primarily on foot;
(d) those with fur and beaks, which move about primarily by flying;
(e) those with feathers and mouths, which move about primarily on foot;
(j) those with feathers and mouths, which move about primarily by flying;
(g) those with feathers and beaks, which move about primarily on foot;
(h) those with feathers and beaks, which move about primarily by flying.
It is not immediately obvious how to assign these (hypothetical) creatures
to categories; any of several schemes (e.g. by coat type, by oral opening type)
would be equally plausible, and none seems particularly reasonable a priori.
Thus, given the total set type of organization, it makes sense that category
assignments should be originally arbitrary. However, it hardly requires
research to demonstrate that the perceived world of objects is not structured
in this manner. Just two of the eight theoretically possible combinations of
attribute values, types a and h (mammals and birds, respectively), comprise
 92 MERVIS & ROSCH

the great majority of existent species in the world that are possible based
on this total set. This correlated attribute structure of the perceived world
has been used as the basis for several programs of research concerning
natural category structure.

Basic Level Categories

Any object may be categorized at each of several different hierarchical
levels. When the levels are related to each other by class inclusion, they
form a taxonomy. Anthropologists working with botanical and zoological
categories ( Berlin et al 1973, C. H. Brown et al 1976) have suggested, on
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

the basis of linguistic and cultural evidence, that one of these levels is more
fundamental than the others. Psychologists (Rosch et a1 1976a) have argued
that the most cognitively efficient, and therefore the most basic level of
by University of Sussex on 10/12/12. For personal use only.

categorization, is that at which the information value of attribute clusters

is maximized. This is the level at which categories maximize within-cate­
gory similarity relative to between-category similarity.
Several studies have investigated hierarchical levels of abstraction. Rosch
et al (1976a) performed attribute analyses for three-level hierarchies of
common concrete objects (e.g. furniture, chair, easy chair) and found the
level corresponding to the level of "chair" to possess the characteristics of
basic level categories. Cantor et al ( 1980) have confirmed that a basic level
also exists for psychiatric category hierarchies. Using somewhat different
measures, Tversky (1977) and Hunn (1976) have found the same level to
be basic.
If this level of categorization is really the most fundamental, one would
expect categories at this level to have special properties. In fact, several such
properties have been identified. Rosch et al ( 1976a) have shown that the
basic level is the most general level at which (a) a person uses similar motor
actions for interacting with category members, (b) category members have
similar overall shapes, and (c) a mental image can reflect the entire cate­
gory. Hunn ( 1975) has argued that the basic level is the only level at which
category membership can be determined by an overall Gestalt perception
without an attribute analysis. Rosch et al ( l976a) have shown that objects
are recognized as members of basic level categories more rapidly than as
members of categories at other levels. In language, the basic level is the one
at which adults spontaneously name objects, whether for adults (Rosch et
a1 1976a) or for young children (R. Brown 1958, 1976; Anglin 1977). Cruse
(1977) has argued that labels for basic level categories are unmarked linguis­
tically-that is, words at this level are used in normal everyday conversa­
tion. In American Sign Language basic level categories are generally
denoted by single signs, while superordinate and subordinate categories are
almost always denoted by multiple sign sequences (Newport & Bellugi
1978).
 CATEGORIZATION OF OBJECTS 93

One of the most pervasive research findings is that basic level categories
are acquired before categories at other hierarchical levels. Rosch et al
(1976a) and Daehler et al ( 1 979) found that young children can solve simple
sorting problems at the basic level before solving them at the superordinate
level. Mervis & Crisafi (198 1), using artificial category hierarchies, found
that 21;2-year-olds were able to sort basic level triads correctly, but could
not sort either superordinate or subordinate level triads. With regard to
language acquisition, Stross ( 1 973) and Dougherty ( 1 978) have both shown
that the first botanical labels that children learn are names for basic level
categories. Other studies have yielded similar results for selected nonbotani­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

cal taxonomies (Rosch et al 1 976a, Anglin 1 977). Taking a historical per­

spective, Berlin (1972) has shown that languages first encode basic level
biological categories, and only later (if at all) encode categories superordi­
by University of Sussex on 10/12/12. For personal use only.

nate or subordinate to the basic level ones.

All of these studies have concerned naturally occurring hierarchies, and
it could be argued that the basic level effects are due to linguistic factors
(e.g. shorter names, greater frequency, learned first) rather than to percep­
tual-cognitive structural factors. Two recent studies, using artificial catego­
ries whose hierarchical structure mirrored the naturally occurring
structure, have suggested that the perceptual-cognitive explanation is more
appropriate.
Murphy & E. E. Smith ( 1 98 1 ) taught subjects to label stimuli at each of
three levels. Order of learning was counterbalanced, and word frequency
and length were controlled. After learning the labels, subjects participated
in a verification task. Response times to verify labels for basic level catego­
ries were significantly shorter than response times to verify labels for subor­
dinate or superordinate level categories. Mervis & Crisafi (1981) controlled
for potential linguistic confounds by never naming the stimuli for their
subjects. One group of subjects was asked to sort the stimuli however it
made sense to them. A second group of subjects was told that a particular
stimulus had been given a certain name, and they should decide which of
the other stimuli should also be given that name. For both tasks, virtually
all of the responses corresponded to the predicted basic level categories.
The principles underlying the determination of which hierarchical level
is basic are expected to be universal. However, for a given domain, the
particular level which is found to be basic may not be universal. This level
can vary as a function of both the cultural significance of the domain and
the level of expertise of the individual (Rosch et al 1 976a, Dougherty 1 978).
These two factors are important because they influence which attributes of
an object are noticed (perhaps constructed) by an observer; psychological
measures of basicness rely on analyses of perceived attribute structures.
Dougherty and Rosch et al both provide examples of the relativity of the
basic level.
 94 MERVIS & ROSCH

Basic Level Categorization as a Basic Process

Flavell & Wellman ( 1976, Flavell 1977) have proposed that memory phe­
nomena be divided into four types: basic processes, knowledge (semantic
memory), strategies, and metamemory. Might basic level categorization be
included as a basic process?
Flavell & Wellman have described two important characteristics of basic
processes. First, a person is not conscious of the actual working of the
process. Second, the process undergoes no significant development (other
than that due to maturation) with age; development is complete by the end
of the sensorimotor period (age Ph to 2 years). They provide three
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

examples of basic processes: 1. the processes by which an object is recog­

nized; 2. the processes of representation underlying recall of absent objects
or events; 3. the process of cueing or associating. They also point out that
by University of Sussex on 10/12/12. For personal use only.

the four types of memory phenomena are not mutually exclusive.

Why should basic level categorization be included as a basic process?
Without any categorization an organism could not interact profitably with
the infinitely distinguishable objects and events it experiences. Therefore,
even infants should be able to categorize. Nevertheless, until recently there
was little motivation to consider infant categorization abilities, since it was
widely believed (see Gelman 1978) that children could not categorize until
they reached the stage of concrete operations (when they are 5 to 7 years
old). However, once simple categorization abilities were demonstrated in
preschool children, research with infants began in earnest. Most of these
studies have taken advantage of an infant's predictable preference for novel
stimuli over familiar ones (L. B. Cohen & Gelber 1975). The studies use the
same general procedure. First, the infant is given several familiarization
trials with different category members. Then he is shown either a novel
member of the same category, a novel member of a different category, or
both at once. If the infant has formed a category, then he should spend
significantly more time looking at a stimulus from a novel category than
from the familiar one. Using this format, G. Ross ( 1 977) demonstrated that
1 2-month-old infants were able to form a variety of basic level categories.
L. B. Cohen demonstrated that much younger infants (30-week-olds) were
able to form the categories "female face" (L. B. Cohen & Strauss 1 979) and
"stuffed animal" (L. B. Cohen & Caputo 1978). Strauss (1979), using sche­
matic faces, demonstrated that lO-month-olds were able to categorize an
average prototype after familiarization with other category members.
Three additional studies have used different techniques. Husaim & L. B.
Cohen (1980), using a discrete trial discrimination learning paradigm,
found that ten-month-olds could form two noncriterially defined categories
(of schematic animals). The infants attended to more than one attribute, and
the same models that predict adult categorization behavior were able to
 CATEGORIZATION OF OBJECTS 95

predict infant behavior. Ricciuti ( 1 965) examined the behavior of 12-, 18-,
and 24-month-olds who were given two types of toys to play with, and
found considerable evidence of categorization abilities in even the 1 2-
month-olds. K. Nelson ( 1 973), using 20-month-olds, has replicated this
result for other basic level categories. In summary, then, there is now
substantial evidence that basic level categorization should be considered a
basic process.

NONEQUIVALENCE OF CATEGORY MEMBERS

Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

In a classical concept formation experiment, any one stimulus which fits the
definition of the concept (possesses the relevant attributes in the correct
combination) is as good an example of the concept as any other. More
by University of Sussex on 10/12/12. For personal use only.

generally, if categories are seen as determinately established by necessary

and sufficient criteria for membership (and if, in addition, the role of ratio­
nality is to abstract out what is essential to a situation while ignoring what
is inessential; see e.g. James 1890a,b), then any member of a category should
be cognitively equivalent qua the category to any other member. However,
there is now a growing amount of empirical evidence that all members are
not equally representative of their category.
The first domain for which nonequivalence was proposed was that of
color. Berlin & Kay ( 1 969) showed that many apparent contradictions
reported in the anthropological literature on color naming could be clarified
by distinguishing focal from nonfocal colors. Focal colors are points in the
color space which speakers of diverse languages agree represent the best
examples of the 1 1 basic color categories. While the number of color terms
in a language and the boundaries of color categories vary widely across
cultures, colors most representative of basic color categories appear to be
universal. These results actually provide cross-linguistic confirmation of
well-established effects in the physiological literature on color naming func­
tions (see Cornsweet 1970). The best exemplars of the four primary colors
correspond to the physiologically determined unique hue points for these
colors (De Valois & Jacobs 1968, Kay & McDaniel 1 978). Standard psycho­
logical variables such as memory accuracy and ease of learning have been
shown to co-vary with representativeness of the color in question (Heider
1972, Rosch 1 974). Findings similar to those with color categories have also
been demonstrated for geometric shape categories (Rosch 1973a,b). For
recent reviews of research on color categories see R. Brown ( 1 976) and
Witkowski & C. H. Brown ( 1 978).
Gradients of representativeness have been found not only for color and
geometric shape categories but also for many common semantic categories
(e.g. "dog," "furniture"). The great majority of psychological studies of
 96 MERVIS & ROSCH

representativeness have focused on such categories. Representativeness is

here defined operationally by means of subjects' ratings of how good an
example an item is of its category (Rosch 1 975b). Consistency in such
ratings has been obtained. Individual subjects agree that some exemplars of
a category are more representative than others, and different subjects con­
sistently choose the same examples as most representative of the category.
The overall scale that is obtained is robust under differing conditions of
instruction and stimulus presentation (Rips et al 1973; Rosch 1973b,
1 975a,b; Rosch & Mervis 1 975; E. E. Smith et al 1 974; Whitfield & Slatter
1979). Two recent studies have found similar gradients for other types of
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

categories: locatives (Erreich & Valian 1 979) and psychiatric classifications

(Cantor et al 1980). Gradients of representativeness for various linguistic
categories have also been widely reported (see e.g. J. Ross 1 972; Fillmore
by University of Sussex on 10/12/12. For personal use only.

1 975, 1 977; Lakoff 1 977; Bowerman 1 978; Bates & MacWinney 1 980;
deVilliers 1 9 80; Maratsos & Chalkley 1 980; Coleman & Kay 1 98 1 ).
Representativeness of items within a category has been shown to affect
virtually all of the major dependent variables used as measures in psycho­
logical research. In this section we consider speed of processing, free pro­
duction of exemplars, natural language use of category terms, asymmetries
in similarity relationships between category exemplars, and learning and
development.
Speed of processing (reaction time) has been extensively investigated in
category verification tasks. Subjects are usually asked to verify statements
of the form "An [exemplar] is a [category name]" as rapidly as possible.
Response times are shorter for verification of the category membership of
representative exemplars than nonrepresentative exemplars; these effects
are robust and appear in a variety, of experimental paradigms (see reviews
in E. E. Smith et al 1 974, E. E. Smith 1 978, Hampton 1979, Danks &
Glucksberg 1980, Kintsch 1980). Rosch et al ( 1976b) have also demon­
strated this effect for three types of artificial categories, where representa­
tiveness was defined by family resemblance, by mean values of attributes,
or by degree of distortion from the prototype (random dot pattern). These
differences in response times are amplified when a prime (prior mention of
the category name) is provided. Priming reduces response times to verify
the category membership of representative exemplars but increases re­
sponse times to verify the membership of nonrepresentative exemplars. This
result has been obtained for colors (Rosch 1 975c), superordinate semantic
categories (Rosch 1975b, MacKenzie & Palermo 198 1), and for the artificial
categories just described (Rosch et al 1 976b).
Order and probability of exemplar production have been investigated
primarily for superordinate semantic categqries. Battig & Montague ( 1969)
asked subjects to list exemplars·.ofeacli 0[-,56 superordinate categories.
Frequency of mention of an exemplar was found to. be significantly corre-
 CATEGORIZATION OF OBJECTS 97

lated with degree of representativeness (Mervis et al 1976). Posnansky

(1973) replicated this result using elementary school children. Both Rosch
et al (1 976b) and Erreich & Valian ( 1 979) found that when subjects were
asked to sketch an exemplar of a particular category, they were most likely
to depict the most representative exemplar.
Natural languages possess mechanisms for coding gradients of represen­
tativeness. For example, languages generally include qualifying terms
("hedges") such as "true" or "technically." Lakoff (1 973) has shown that
a given hedge is applicable to only a subset of category exemplars; this
subset is determined by degree of representativeness. For instance, it is
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

acceptable to say "A sparrow is a true bird," but not "A penguin is a true
bird." Correspondingly, the sentence "A penguin is technically a bird" is
acceptable, but "A sparrow is technically a bird" is not. Similarly, Rosch
by University of Sussex on 10/12/12. For personal use only.

(1 975a) has shown that when subjects are given sentence frames such as "[x]
is virtually [y)," they reliably place the more representative member of a
pair in the referent (y) slot. In addition, representativeness ratings for
members of superordinate categories predict the extent to which the mem­
ber term is substitutable for the superordinate word in sentences (Rosch
1 977). Finally, Newport & Bellugi (1 978) have shown that in American
Sign Language, when superordinate terms are denoted by a short list of
exemplars only the more representative exemplars may be used.
Asymmetry in similarity ratings between members that vary in represen­
tativeness is another way in which members of a category fail to be equiva­
lent. Tversky & Gati ( 1 978) and Rosch ( 1975a) have shown that less
representative exemplars are often considered more similar to more repre­
sentative exemplars than vice versa. For example, subjects felt that Mexico
was more similar to the United States than the United States was to Mexico.
This phenomenon helps to explain the asymmetries which Whitten et al
( 1979) found in similarity ratings of pairs of "synonyms." It also helps to
explain Keller & Kellas's (1978) finding that release from proactive inhibi­
tion is significantly greater if the change is from typical to atypical members
of a category than if the change is from atypical to typical members. In
addition, asymmetry in similarity ratings has implications for inductive
reasoning. Rips (1975) found that new information about a category mem­
ber was generalized asymmetrically; for example, when told that the robins
on an island had a disease, subjects were more likely to decide that ducks
would catch it than that robins would catch a disease which the ducks
had.
In the learning and development of categories, representativeness appears
to be a major variable. Representativeness gradients have two basic implica­
tions for category acquisition (Mervis & Pani 1 980). The first implication
is that category membership is established (for the set of exemplars to which
a person has been exposed) first for the most representative exemplars and
 98 MERVIS & ROSCH

last for the least representative exemplars. One of the most robust findings
from research using statistically generated categories is that correct classifi­
cation of novel exemplars is strongly negatively correlated with degree of
distortion of the exemplar from the prototype pattern. This result has been
obtained using both random dot pattern categories (e.g. Posner & Keele
1968, Homa et al 1973, Homa & Vosburgh 1 976) and random polygon
categories (e.g. Aiken & Williams 1973, Williams et al 1977). Similarly,
when subjects are asked to indicate which of a series of categorically related
stimuli have been seen previously, percentage of false recognition responses
and degree of confidence that the (novel) pattern has been seen previously
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

are both negatively correlated with degree of distortion from the prototype
(e.g. Franks & Bransford 1 97 1 ; Neumann 1 974, 1 977; Posnansky & Neu­
mann 1 976). When subjects are given explicit feedback concerning the
by University of Sussex on 10/12/12. For personal use only.

correctness of their classifications, categories consisting of low distortions

are learned significantly faster than categories consisting of either high
distortions or both low and high distortions (e.g. Posner et al 1 967, Homa
& Vosburgh 1916). For categories including both low and high distortions,
the low distortions are learned first (Mirman 1 978). Similarly, Rosch
( 1 973a,b) found that focal (representative) colors and forms were learned
more rapidly than nonfocal colors and forms by persons whose language did
not contain explicit labels for these categories. In a study of 5-year-olds
learning artificial categories modeled after natural categories, Mervis &
Pani (1 980) found that more representative exemplars were learned first; in
this study, no feedback was provided during learning.
The developmental research relevant to this issue has concentrated pri­
marily on the acquisition of superordinate semantic categories. K. E. Nel­
son & K. Nelson (1 978) have argued that as children learn about a category,
their criteria for assigning an object to that category shift back and forth
between generous and conservative, until finally the adult criteria (which
themselves vary according to cognitive style; see Kogan 1971) are used. This
pendulum theory predicts that category membership of representative ex­
emplars should be firmly established at a young age, while membership of
less representative exemplars will vascillate. Although no single study has
considered a wide enough age range to test the theory conclusively, it
appears, based on the combined results from several sorting studies using
children of different ages (Saltz et al 1 972, Neimark 1 974, Anglin 1 977),
that the theory may be correct. Two studies of children's production of
category exemplars (K. Nelson 1974, Rosner & Hayes 1 977) provided
additional support. Research with basic color categories (Mervis et a1 1 975)
and with basic object categories (Saltz et al 1 977) support the finding of the
superordinate sorting studies.
The second major implication of representativeness is that categories are
learned more easily and more accurately if initial exposure to the category
 CATEGORIZATION OF OBJECTS 99

is through only representative exemplars. Two studies have shown that

initial exposure to only representative exemplars is more effective than
initial exposure to only nonrepresentative exemplars; the stimuli were dot
patterns (Mirman 1978) and multimodal artificial stimuli with natural cate­
gory structure (Mervis & Pani 1980).
Results are somewhat more equivocal when initial exposure to both
representative and nonrepresentative examples is compared with initial
exposure to only representative examples. Two studies (Homa & Vosburgh
1976, Goldman & Homa 1977) found that for categories with certain char­
acteristics initial exposure to the full range of category membership was not
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

worse than exposure to good examples only. However, three other studies
found training on good examples superior to training on a range of examples
(Mervis & Pani 1980, Hupp & Mervis 198 1 , Mervis & Mirman 198 1).
by University of Sussex on 10/12/12. For personal use only.

What might make some objects more representative of their category

than others? It may be useful to consider once more the previous section
in which we discussed clusters of correlated attributes. These correlations
are not perfect-for example, in our hypothetical set of creatures (p. 9 1),
besides the two main correlated clusters (birds and mammals), there are
many types of flightless birds, a few flying mammals (bats), and the duck­
billed platypus. It has been argued that the logic of attribute structures
associated with gradients of representativeness within categories is parallel
to the logic (described in the previous section) that predicts which level in
a taxonomy will be the basic level (Rosch 1 978, Mervis 1980).
Rosch & Mervis (1975) have shown that category members differ in the
extent to which they share attributes with other category members. They
call this variable family resemblance (after Wittgenstein 1 953). Items which
have the highest family resemblance scores are those with the most shared
attributes. Rosch & Mervis (1975) have also shown that the exemplars with
the highest family resemblance scores are those which share few (if any)
attributes with members of related categories. In other words, given the
nature of real-world attribute clusters, the items that have most attributes
in common with other members of their own category also have fewest
attributes in common with related contrast categories. Both family resem­
blance and dissimilarity from contrast categories are highly correlated with
ratings of representativeness for superordinate and basic level natural cate­
gories and for artificial categories (Rosch & Mervis 1 975). The within­
category correlations have been confirmed using different measures
(Neumann 1977, Tversky 1 977, Tversky & Gati 1978). Studies with various
artificial categories have shown that when within-category similarity and
between-category dissimilarity are dissociated, either factor is sufficient to
produce a representativeness gradient (Rosch et al 1 976b, E. E. Smith &
Balzano 1 977). (Note that the family resemblance idea does not indicate
that category members must have no attributes common to all members-
 100 MERVIS & ROSCH

in fact, insofar as category members are the same higher level "sort of
thing" they will share such higher level, and therefore necessary, attributes
as "animate" or "solid object"; see Keil 1 979, E. E. Smith & Medin 1 981.)
We can now see how a family resemblance structure of categories might
make sense of the findings that the most representative members of catego­
ries are established first as category members and are the most useful basis
for learning categories. Because basic level categories maximize within­
category similarity relative to between-category similarity, it is reason­
able that they were found to be learned first-before categories subordinate
and superordinate to them. Correspondingly, the most representative exem­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

plars of a category have maximal within-category and minimal between­

category similarity. Therefore, category membership is most obvious for the
highly representative exemplars, and generalization based on similarity to
by University of Sussex on 10/12/12. For personal use only.

these will be the most accurate.

The above findings on the nonequivalence of natural category members
have been mirrored in research on social phenomena. Relying on the corre­
lated attribute cluster as a basis for assigning category membership serves
to make the world seem more orderly than it really is (RoSCh 1978, Mervis
1980). Lippmann (1922) has argued that social stereotypes serve exactly this
purpose. Kahneman & Tversky (1973) have shown that when subjects are
asked to predict a person's occupation based on a description of the person,
the person is assigned to the occupation for which the best match between
personal description and occupational stereotype is obtained. C. E. Cohen
( 1 976) found that occupational and role stereotypes are used as a basis for
inferring a person's characteristics. Cantor & Mischel (1979) showed the
usefulness of a family resemblance stereotype notion for determining attri­
bution of personality trait categories such as "introvert" and "extrovert."
McCauley & Stitt (1978) have presented a Bayesian method for determining
which attributes should be included in a stereotype; their method is remi­
niscent of family resemblances. In fact, McCauley & Stitt have argued that
their method is also applicable to concrete object representations, since
stereotypes of groups and representations of object categories serve the same
function. For reviews of research concerning stereotypes, see Brigham
(1971) and McCauley et al ( 1 980).

INDETERMINACY OF CATEGORY MEMBERSHIP

AND REPRESENTATION

That category members vary in degree of representativeness is essentially

a set of empirical findings. What does it imply about the nature of catego­
ries? There are two separable issues: the relatively empirical question of
whether the boundaries of categories are determinate (well defined) or
 CATEGORIZATION OF OBJECTS 101

fuzzy, and the theoretical issue of how well defined one wants to consider
the category "representation" itself.
Two experimental approaches have been used to demonstrate that cate­
gory boundaries are not well defined. The first involves demonstrations that
there are between-subject disagreements concerning which categories cer­
tain (poor) exemplars belong to. Berlin & Kay ( 1 969) found substantial
disagreement among subjects concerning the location of color category
boundaries, even when native language was controlled. This result has been
replicated by Labov ( 1973), using schematic drawings of cups, by McClos­
key & Glucksberg (1978), using superordinate semantic categories, by Can­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

tor et al ( 1 980), using subordinate psychiatric categories, and by Kempton

(1978) using dri�king vessels in a cross-cultural setting.
The demonstration of between-subject disagreement is suggestive, but
by University of Sussex on 10/12/12. For personal use only.

this disagreement could possibly be an artifact stemming from the combin­

ing of data from different subjects, each of whose categories had a different
well-defined cutoff. Therefore, a demonstration of within-subject disagree­
ment is also important. Berlin & Kay ( 1 969) noted substantial within­
subject disagreement across testing sessions concerning color category
boundaries. McCloskey & Glucksberg (1978) found the same result for
multiple within-subject judgments concerning superordinate category as­
signments of potential poor exemplars. They also demonstrated that their
results could not be explained simply by reference to polysemous superordi­
nate category labels.
Another empirical consideration points to the reasonableness of nondefi­
nite boundaries of categories. Poorer members of categories are likely to
contain attributes from the correlated attribute clusters of other categories
(see e.g. Rosch & Mervis 1 975). Sokal (1974) has provided an elegant
demonstration of this point for biological categories (see also Simpson 1 961,
Sneath & Sokal 1973).
The controversy over the determinacy of categories, however, extends
beyond empirical evidence. In present cognitive psychology it has become
almost obligatory to explain and model phenomena in terms of cognitive
representations and processes which act on them. If one believes that cate­
gories consist of determinate necessary and sufficient criteria, one can de­
velop a model which attempts to explain representativeness and indeter­
minate boundary effects by means of processes operating on a determinate
representation. For example, Wanner ( 1 979) reported the finding that
mathematical concepts such as "odd number" show the same representa­
tiveness effects described in the Nonequivalence section. He argued that
since mathematical concepts are the archetype of concepts "true by defini­
tion alone," one may interpret these results to imply that the "real mean­
ing" of a concept consists of a criterial definition; representativeness effects
would be produced by separate processing heuristics. [It is interesting to
 102 MERVIS & ROSCH

note in this context that a major contemporary school of mathematics (the

constructivist school) does not consider mathematical concepts to be either
true by definition alone or necessarily criterial; see Calder 1979.] Glass &
Holyoak ( 1975) have developed a model in which concepts are represented
by semantic markers (Katz 1972) which contain the essential features of the
concept. These markers are nodes in semantic memory connected by path­
ways; representativeness effects are modeled by means of variation in length
and directness of these pathways. Another strategy is used in the fuzzy set
(Zadeh 1965) model of Caramazza ( 1979). In this model noun concepts are
represented by determinate specification of defining attributes, but the at­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

tributes (which are presumed not to be concepts) are treated as fuzzy. E.

E. Smith & Medin (198 1) provide a general characterization of such models.
They argue that such models cannot account for a variety of empirical
by University of Sussex on 10/12/12. For personal use only.

findings.
Other models incorporate greater indeterminacy into the representation
itself. In the two-stage model of E. E. Smith et al (1 974) and the single-stage
models of McClosky & Glucksberg (1979) and Hampton ( 1979), concepts
are represented by sets of weighted non-necessary features. Processing deci­
sions about category membership are made on a probabilistic basis. Collins
& Loftus's (1975) model also includes a probabilistic decision process, but
it is allowed to operate on more varied types of structural information than
simply attributes.
The issue of determinacy has been approached in a slightly different
fashion in research on context effects. Both Collins & Loftus (1975) and
McClosky & Glucksberg (1 979) invoke Bayesian inference procedures spe­
cifically to deal with context; however, all of the probabilistic models have
an unchanging (and, in that sense, determinate) representation of some sort.
Context has been employed to criticize just this sort of assumption. In the
psychological literature, changes in meaning, comprehension, or memory
of particular terms as a function of differing contexts have been used to
question the adequacy of semantic memory models (see e.g. Barclay et al
1 974, R. C. Anderson & Ortony 1975, R. C. Anderson et al 1 976, Potter
& Faulconer 1979). These studies tend to be primarily critical rather than
to offer formulations of their own. In addition, context has been used as the
basis of more far reaching criticisms of determinate views; for authors in
the hermeneutic tradition, context becomes the basis of arguments against
representations and other noninteractive accounts of meaning (see for ex­
ample the papers in Rabinow & Sullivan 1979).

THE NATURE OF ABSTRACTION

In the classic concept formation paradigm, a concept is an abstraction

consisting of a set of defining features and the relationship between them.
 CATEGORIZATION OF OBJECTS 103

Contemporary views have argued both that concepts may be conceived as

less abstract or must be conceived as more abstract than this formulation.
There are, of course, also accounts that posit intermediate levels of abstrac­
tion.
On the one hand, it has been asserted that we need only memory for
individual exemplars in order to account for categorization. In a sense, from
the time of Pavlov, "strength theory" models of stimulus generalization
have been of this type (see Riley & Lamb 1 979). In human categorization
research, Reber (1976) and Brooks (1978) have demonstrated cases where
learning of instances can occur without the learning of rules or abstractions;
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

in fact, where telling a subject the rule may retard performance. If only
specific exemplars are stored in memory, categorization of novel items could
occur by matching the new instance to the most similar item in memory.
by University of Sussex on 10/12/12. For personal use only.

Formal exemplar ("nearest neighbor") models which incorporate this type

of processing have been tested against various abstraction models; in all
cases, exemplar models were inferior to the abstraction models (Reed 1972,
Hyman & Frost 1975, Hayes-Roth & Hayes-Roth 1 977). In addition, in
order to account for hierarchical relationships, exemplar models must re­
quire that the names of all possible superordinate categories be stored with
each exemplar. Meyer (1 970) has discussed problems with this approach.
Given the state of present cognitive modeling, exemplar models are also
theoretically anomalous for two reasons. First, if beyond learning items
cognitive representations of them are required, instances must either be
coded by first order isomorphism (the representation of "green square" as
green and square; Shepard & Chipman 1970, Palmer 1978) or undergo some
kind of abstractive process. Second, nearest neighbor models require an
account of similarity judgments; presently such accounts all involve ab­
stractions (Tversky 1977).
Virtually all models of categorization involve abstraction-that is, ways
in which the cognitive system acts "creatively" on input during learning of
categories and uses the resultant categorical information to classify novel
items. The creativity is of two types: determining which elements of a
situation are "essential" and which irrelevant; and the creation of new
higher order information which was not given in any particular exemplar.
In the classical concept formation paradigm, abstraction of essential ele­
ments is involved in learning which attributes are relevant, and creation of
higher order information is involved in learning the logical relationship
between these attributes. Any model that includes representation offeatures
(whether defining or characteristic) posits creativity of the first type. The
type of novel higher order information generated varies for different models.
The most minimal computation is required by prototype models based on
central tendency: e.g. on means (Posner et al 1 967, Posner & Keele 1 968,
Reed 1 972, Reed & Friedman 1973), modes (Neumann 1 974, 1977; Hayes-
 104 MERVIS & ROSCH

Roth & Hayes-Roth 1977) or ideal values based on perceptual characteris­

tics (Kay & McDaniel 1978, Oden & Massaro 1978). Some models posit
that category representations include both summary information and exem­
plar information (Medin & Schaffer 1 978, Smith & Medin 198 1 ). For their
summary information, these models require abstraction of "essential" ele­
ments to generate features and the creation of new information in the form
of weights for the features (computed, in general, on the basis of the useful­
ness of the feature for determining category membership).
Abstraction models involving simple attributes, rules, or prototypes
(such as any of the above) are criticized by those who feel that higher order
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

abstractions and general knowledge more extensive than that of individual

categories are required in any account of categorization. For example,
Pittenger & Shaw (1975, Pittenger et al 1979) argue that higher order
by University of Sussex on 10/12/12. For personal use only.

knowledge about transformations serves as the perceptual invariant under­

lying certain types of categorization. In the tradition of constructive mem­
ory research (e.g. Bartlett 1932, Bransford & McCarrell 1 974), categories
are treated as part of very general schemas. Finally, large scale computer
models (Collins & Loftus 1975, Schank & Abelson 1 977, Winograd 1 972)
treat categories and categorization processes as inseparable from world
knowledge and the inference processes used in such knowledge.

DECOMPOSABILITY OF CATEGORIES INTO

ELEMENTS

Virtually all accounts of the representation and processing of categories

assume that categories are decomposable into more elementary qualities.
This is not surprising: as Dreyfus (1979) has pointed out, since the time of
Plato one of the major aspects of what has been meant by an explanation
has been the decomposition of the thing to be explained into its elements.
In psychology, however, arguments against the indiscriminate use of expla­
nation-by-decomposition have been with the field since its inception (see e.g.
James 1 890a on the unitary nature of a single complex thought). The issue
of decomposability was the focus of the debate in the early part of this
century between the structuralists, who saw all experience as built out of
primitive meaningless sensations, and the GestaItists, who emphasized irre­
ducible emergent properties of wholes (see Boring 1 950).
At present, while decomposition is the unmarked assumption in model
building, the possible need for less analytic factors is periodically acknowl­
edged. For example, in the pattern recognition literature, analysis into
features and holistic matching to a template are generally presented as the
two major types of alternative models (Reed 1 973), although for good
reasons templates are usually treated as straw men (see Palmer 1 978).
 CATEGORIZATION OF OBJECTS 105

Various empirical developments have brought the nature and role of

decomposition into current debate. First, categorization has been investi­
gated for types of stimuli that do not have obvious elements at a cognitive
level. The most notable of these are color (Rosch 1 973b, 1974) and overall
configuration (e.g. Attneave 1 957, Posner 1969, Lockhead 1 972-but see
also Barresi et a1 1 975, Homa & Vosburgh 1 976). It is not surprising that
such stimuli do not strike us as obviously decomposible, since they are
themselves normally treated as attributes, i.e. the qualities into which more
complex objects are decomposed. There has also been considerable dis­
agreement as to whether faces should be considered special holistically
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

perceived objects (Hochberg & Galper 1 967, Yin 1 969, Rock 1 973, Brad­
shaw 1 976). A second development which has offered the opportunity to
view categories as wholes is the possibility for spatial representation of
by University of Sussex on 10/12/12. For personal use only.

within- and between-category structures through techniques such as mul­

tidimensional scaling. For example, Hutchinson & Lockhead (1977) have
argued that categories can best be conceived as unanalyzed points in metric
multidimensional space. A third trend has been use of the concept of a
prototype and the facts of gradients of representativeness to suggest holistic
processing (e.g. Rosch 1973b, Dreyfus 1 979).
The great majority of arguments over decomposition concern specifying
the level of abstraction at which a particular kind of decomposition can or
cannot be said to occur. While most categorization models include decom­
position, it is never to the point of infinite regress. Some elements are
included as the primitives, although usually by default, rather than by
explicit labeling as primitives.
Because some elements are not decomposed, many accounts of catego­
rization include an explicit holistic component. For example, this can be
introduced by means of a (relatively) holistic processing stage (E. E. Smith
et al 1 974). Another possibility is that a given level of abstraction may be
a basic and (potentially) holistically perceived level, even if other levels
require more analytic mechanisms (Rosch et al 1 976a). Perceptual process­
ing of figures (such as a large letter constructed of smaller letters) has been
shown to proceed from global to local analysis under some stimulus condi­
tions (Navon 1 977) but to proceed from local to global under others (e.g.
Kinchla & Wolf 1979). In the perception models of Palmer (1975) and
Winston ( 1975), the decomposition of a visual scene is viewed as a hierar­
chical network of subscenes, and it is claimed that higher-order properties
are processed first, followed by lower-order properties. Depending on the
circumstances, however, a given aspect of a scene might be either the
higher- or the lower-order property. The controversy over lexical decompo­
sition in linguistics and artificial intelligence (does "kill" really mean "cause
to die") can be seen as largely an issue of which linguistic level to consider
a whole and which to consider the elements (Fodor 1 970; McCawley 1971,
 106 MERVIS & ROSCH

1 978; Schank 1973; Fodor et al 1980). The recent suggestion that the
capacity to match figures holistically may be a cognitive strategy which
shows individual differences (Cooper 1 976, Cooper & Podgorny 1 976) is
intriguing; for contrasting work on individual differences see Day ( 1 976).
Two important points have been made in this discussion of decomposabil­
ity. First, although the tendency in cognitive models is to decompose almost
automatically, the evidence of holistic processing of some stimuli or at some
stages suggests that we be more thoughtful about decompositional models.
Second, findings concerning decomposition appear to be dependent on the
level under consideration. Some general principles of decomposition are
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

needed.

THE NATURE OF ATTRIBUTES

by University of Sussex on 10/12/12. For personal use only.

In the British empiricist account, attributes correspond to elementary sen­

sations. However, in modern cognitive psychology, almost anything has
been used as an attribute at one time or another. This produces some
anomalies, particularly in the use of parts, relations, and functions as at­
tributes (Rosch 1 978). Indeed, as pointed out in the previous section, what
is considered a category and what are called its attributes depend on the
level one is describing; the same item (e.g. "red" or "circular") can be what
is to be explained (category) or what is referred to as part of the explanation
(attribute). Appropriately, therefore, there have recently been a number of
discussions of the nature of attributes and the manner in which they com­
bine.
The first controversy in the field involves use of features vs use of dimen­
sions in the representation of categories. Features generally designate quali­
tative properties (e.g. legs, wooden, you sit on it) and so need not be
applicable to all objects in the same domain. Large numbers of features may
be included in a single representation. There are many different types of
feature representations, such as feature lists and structural descriptions; for
an extensive discussion see Palmer ( 1 978). Explanations using features are
overtly decompositional.
In contrast, dimensions are usually employed to describe quantitative
properties (e.g. size); therefore, every object in a given domain is assigned
some value on each of the dimensions used to describe the domain. An ideal
dimensional representation includes only a small number of dimensions.
There are two different types of dimensional representation, metric and
nonmetric; for a discussion of their use in categorization models see E. E.
Smith & Medin ( 1 98 1). Dimensional descriptions often use a spatial meta­
phor (encouraged by the availability of multidimensional scaling tech­
niques) and thus may appear to be relatively holistic representations of
categories.
 CATEGORIZATION OF OBJECTS 1 07

However, as Palmer (1978) and E. E. Smith & Medin (1981) have pointed
out, the difference between features and dimensions may apply more to the
surface form of the representation than to the underlying information that
is represented. It can be shown that features may be extended to handle
quantitative properties, and dimensions may be extended to handle most
(but perhaps not all; see Beals et a1 1968) qualitative properties. (But note
that missing values communicate different information for features vs for
dimensions; see Garner 1 978a,b.) Features, transformations, and nonmetric
dimensions can be integrated reasonably into the same representation.
One of the first psychologists to question the nature of attributes (whether
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

features or dimensions) was Garner (1970, 1974). Garner was concerned

not with the form by which attributes should be represented in cognitive
models, but rather with a proposed difference in kind among perceptual
by University of Sussex on 10/12/12. For personal use only.

attributes. This difference is concerned with how types of attributes com­

bine with each other. Garner distinguished two types of attribute combina­
tions: those that are separable (e.g. form and size) and those that are integral
(e.g. brightness and saturation). Attribute combinations are considered
separable if they are perceived in terms of their separate attributes; simi­
larity is therefore judged by comparing the relevant stimuli with regard to
their values on each of the component attributes. Attribute combinations
are considered integral if the two attributes are not treated separately, that
is, if a change in one attribute appears to produce a stimulus which is
different as a whole rather than different for that one attribute. For integral
attributes, similarity is judged holistically, according to how much the
relevant stimuli are alike.
This distinction between types of attributes appeared temporarily to have
major developmental implications. For adults, certain combinations of at­
tributes are separable, while others are integral; however, young children
appeared to treat virtually all attribute combinations as though they were
integral (see e.g. Shepp & Swartz 1 976, Shepp 1 978). Clarification of the
developmental data results from a further consideration of the nature of
integral combinations. It has been suggested (e.g. Lockhead 1972, Garner
1974) that there are actually two types of integral combinations: those
which seem to be mandatorily perceived holistically, and those which peo­
ple prefer to process holistically, but which can also be processed dimen­
sionally if such processing is advantageous. Recently, there have been
several demonstrations (e.g. L. B. Smith & Kemler 1 977, 1 978; Kemler &
L. B. Smith 1978) that young children treat those attribute combinations
which are separable for adults as though they corresponded to the second
type of integral combination. Thus, children are in fact able to perceive each
of the attributes and to treat them as though they were separable; young
children simply prefer to attend to holistic relationships. Young children
are especially likely to consider separable dimensions separately if the re-
 108 MERVIS & ROSCH

quired task is conceptual rather than perceptual (Kemler & L. B. Smith

1 979).
The question arises as to what makes some attributes combine in an
integral fashion and others in a separable fashion. Aside from some initial
speculation (Gamer 1 974), this question has not been pursued. It may be
pointed out, however, that integral and separable attributes appear to be at
different levels of abstraction. When stimulus dimensions are considered at
the level at which one normally calls something an attribute (e.g. at the level
of colors and forms), the attributes are separable. Integral attributes are
further decompositions of that level (e.g. hue and brightness are further
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

decompositions of color). Thus separability and integrality of attributes,

like other issues in decomposition, appear to depend on the level of abstrac­
tion considered.
by University of Sussex on 10/12/12. For personal use only.

Integrality and separability of attributes may be considered part of the

general issue of information integration. Even given separable attributes,
there remains the question of how to model the way in which they combine.
The main choice in formal models has been between additive and multi­
plicative. Additive models treat attributes as though they were independent.
These models appear to work best when relevant information is presented
sequentially and correlations between attributes are absent or not apparent.
Multiplicative models treat attributes as though they were nonindependent,
and therefore work best when relevant information is presented simulta­
neously and correlations between attributes are apparent. (For a review of
information integration see N. H. Anderson 1 974.) In categorization re­
search, additive and multiplicative models may make very similar predic­
tions for real-world categories, since the attributes which are correlated are
generally also most frequent.
Let us return to the original role of attributes in categorization theory.
In empiricist philosophy, attributes were used to connect concepts to the
real world; that is, to connect the meaning (intension) of a concept with the
objects (extension) which fit that meaning. In the psychology of categoriza­
tion, attributes are often used for this purpose. These attributes are gener­
ally of four types: parts, physical characteristics such as color and shape,
relational concepts such as taller, and functional concepts. However, these
types of attributes as represented in categorization models are all categories
themselves; therefore, they can themselves be examined as a categorization
problem (see Rosch 1 979).
Note that one major psychological theory takes a different approach to
the origin of attributes. A constructivist approach to logical classification
(Piaget 1 970, 1 972) takes as its unit of analysis the interaction of persons
and objects. Attributes are developed out of this interaction. Perhaps the
closest analogy to such an approach in non-Piagetian cognitive psychology
is the current interest in modeling categories by means of procedures. At
 CATEGORIZATION OF OBJECTS 109

present, this work is largely confined to formal systems (Miller & Johnson­
Laird 1 976) and to artificial intelligence (see Winograd 1975).

SUMMARY

New trends in categorization research have brought into investigation and

debate some of the major issues in conception and learning whose solution
had been unquestioned in earlier approaches. Empirical findings have estab­
lished that: (a) categories are internally structured by gradients of repres en­
tativeness; (b) category boundaries are not necessarily definite; (c) there is
a close relation between attribute clusters and the structure and formation
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

of categories. This appears to be a particularly promising approach for

future research.
These findings challenge determinate definitions of categories and pro­
by University of Sussex on 10/12/12. For personal use only.

vide constraints on alternative views. Other issues that research and theory
in the modeling of categorization have brought into focus are the nature of
the abstractive process, the question of decomposition of categories into
elements, and the nature of the attributes into which categories are often
decomposed. In short, current research on categories could be said to
represent a kind of experimental epistemology.

ACKNOWLEDGMENTS

We would like to thank Doug Medin, Elissa Newport, Emilie Roth, and
especially John Pani for their thoughtful criticisms of previous drafts of this
paper. Preparation of the manuscript was supported by grant No. BNS79-
1 5 1 20 from the National Science Foundation and by grant No. 6-400-
76-01 1 6 from the National Institutes of Education, both to the University
of Illinois, Urbana-Champaign; and by grant No. I RO I MH243 16-03 from
the National Institutes of Mental Health to the University of California,
Berkeley.

Literature Cited

Aiken, L. S., Williams, T. M. 1973. A devel­ terms. J. Verb. Learn. Verb. Behav.
opmental study of schematic concept 1 5 :667-79
formation. Dev. Psycho!. 8 : 1 62-67 Anglin, J. M. 1977. Word, Object and Con­
Anderson, N. H. 1974. Algebraic models in ceptual Development. New York: Nor­
perception. In Handbook of Perception, ton. 302 pp.
ed. E. C. Carterctte, M. P. Fried­ Attneave, F. 1957. Transfer of experience
man, 2:21 5-99. New York: Academic. with a class-schema to identification­
556 pp. learning of patterns and shapes. J. Exp.
Anderson, R. C., Ortony, A. 1975. On put­ Psychol. 54:81-88
ting apples into bottles-A problem of Barclay, J. R., Bransford, J. D., Franks, J. J.,
polysemy. Cognit. Psycho!. 7:1 67-80 McCarrell, N. S., Nitsch, K. 1974.
Anderson, R. C., Pichert, J. W., Goetz, E. T., Comprehension and semantic flexibil­
SchaJlert, D. L., Stevens, K. V., Trollip, ity. J. Verb. Learn. Verb. Behav. 13:
S. R. 1 976. Instantiation of general 471-81
 1 10 MERVIS & ROSCH

Barresi, 1., Robbins, D., Shain, K. 1975. Role B. B. Lloyd, pp. 1 69-2 1 5. Hillsdale, NI:
of distinctive features in the abstraction Erlbaum. 328 pp.
of related concepts. 1. Exp. Psychol. ­ Brown, C. H., Kolar, J., Torrey, B. I.,
Hum. Learn. Mem. 104:360--68 Truong-Quang, T., Volkman, P. 1976.
Bartlett, F. C. 1932. Remembering: A Study Some general principles of biological
in Experimental and Social Psychology. and non-biological folk classification.
Cambridge: Cambridge Univ. Press. Am. Ethnol. 3:73-85
3 1 7 pp. Brown, R. 1958. How shall a thing be called?
Bates, E., MacWinney, B. 1980. Functional­ Psychol. Rev. 65: 14-21
ist approaches to grammar. In Lan­ Brown, R. 1976. Reference: In memorial
guage Acquisition: The State of the tribute to Eric Lenneberg. Cognition
Art. ed. L. Gleitman, E. Wanner. 4: 1 25-53
Cambridge: Cambridge Univ. Press. Bruner, I. S., Goodnow, I. I., Austin, J. G.
In press 1956. A Study of Thinking. New York:
Battig, W. F., Montague, W. E. 1969. Cate­ Wiley. 330 pp.
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

gory norms for verbal items in 56 cate­ Calder, A. 1979. Constructive mathematics.
gories: A replication and extension of Sci. Am. 24 1 : 1 46-69
the Connecticut category norms. J. Cantor, N., Mischel, W. 1979. Prototypes in
Exp. Psychol. Monogr. 80: 1--46 person perception. In Adv. Exp. Soc.
by University of Sussex on 10/12/12. For personal use only.

Beals, R., Krantz, D., Tversky, A. 1968. Psychol. 1 2:3-52

Foundations of multidimensional scal­ Cantor, N., Smith, E. E., French, R. D., Mez­
ing. Psychol. Rev. 75: 1 27--42 zich, I. 1980. Psychiatric diagnosis as
Berlin, B. 1972. Speculations on the growth prototype categorization. J. Abnorm.
of ethnobotanical nomenclature. Lang. Psychol 89: 1 8 1-93
Soc. 1 :5 1-86 Caramazza, A. 1979. Are concepts ill-defined
Berlin, B., Breedlove, D. E., Raven, P. H. or vague? Presented at Conf. Word &
1973. General principles of classifica­ Concept, Stanford, Calif.
tion and nomenclature in folk biology. Cohen, C. E. 1976. Cognitive basis ofstereoty­
Am. Anthropol 75:2 1 4--42 ping: An information-processing ap­
Berlin, B., Kay, P. 1969. Basic Color Terms: proach to social perception. PhD disser­
Their Universality and Evolution. Ber­ tation. Univ. Calif., San Diego
keley: Univ. Calif. Press. 1 78 pp. Cohen, L. B., Caputo, N. F. 1978. Instructing
Boring, E. G. 1950. A History of Experimen­ infants to respond to perceptual catego­
tal Psychology. New York: Appleton­ ries. Presented at Midwestern Psycho!.
Century-Crofts. 777 pp. 2nd ed. Assoc. Conv., Chicago
Bourne, L. E. Jr., Dominowski, R. L., Loftus, Cohen, L. B., Gelber, E. R. 1975. Infant vi­
E. F. 1979. Cognitive Processes. Engle­ sual memory. In Infant Perception:
wood Cliffs, NI: Prentice-Hall. 408 pp. From Sensation to Cognition: Basic Vi­
Bowerman, M. 1978. The acquisition of word sual Processes. ed. L. B. Cohen, P.
meaning: An investigation into some Salapatek, 1 :347-403. New York: Aca­
current conflicts. In Development of demic. 426 pp.
Communication: Social and Pragmatic Cohen, L. B., Strauss, M. S. 1979. Concept
Factors in Language Acquisition, ed. N. acquisition in the human infant. Child
Waterson, C. Snow, pp. 263-87. New Dev. 50:419-24
York: Wiley. 498 pp. Coleman, L., Kay, P. 1 9 8 1 . Prototype seman­
Bradshaw, M. 1976. An investigation of tics: The English word lie. Language,
stimulus integrality in the perception of Vol. 57. In press
schematic faces. Tech. Rep. 82. Dep. Collins, A. M., Loftus, E. F. 1975. A spread­
PsychoL. Johns Hopkins Univ. ing activation theory of semantic pro­
Bransford, I. D., McCarrell, N. S. 1 974. A cessing. Psychol. Rev. 82:407-28
sketch of a cognitive approach to com­ Cooper, L. A. 1976. Individual differences in
prehension: Some thoughts about visual comparison processes. Percept.
understanding what it means to com­ Psychophys. 19:433-44
prehend. In Cognition and the Symbolic Cooper, L. A., Podgorny, P. 1976. Mental
Processes. ed. W. B. Weimer, D. S. Pa­ transformations and visual comparison
lermo, pp. 1 89-229. New York: Wiley. processes: Effects of complexity and
450 pp. similarity. J. Exp. Psychol. -Hum. Per­
Brigham, I. C. 1 9 7 1 . Ethnic stereotypes. Psy­ cept. Perform. 2:503-14
chol. Bull 76: 1 5-38 Cornsweet, T. N. 1 970. Visual Perception.
Brooks, L. 1978. Nonanalytic concept forma­ New York: Academic. 475 pp.
tion and memory for instances. In Cog­ Cruse, D. A. 1977. The pragmatics of lexical
nition and Categorization. ed. E. Rosch, specificity. 1. Ling. 1 3 : 1 53-64
 CATEGORIZATION OF OBJECTS 111

Daehler, M. W., Lonardo, R., Bukatko, D. motion and Structure. New York: Wi­
1979. Matching and equivalence judg­ ley. 203 pp.
ments in very young children. Child Garner, W. R. 1978a. Aspects of a stimulus:
Dev. 50: 1 70-79 Features, dimensions, and configura­
Danks, J. H., Glucksberg, S. 1980. Experi­ tions. See Brooks 1978, pp. 99-133
mental psycholinguistics. Ann. Rev. Garner, W. R. 1 978b. Selective attention to
Psychol. 3 1 :391-417 attributes and to stimuli. J. Exp. Psy­
Day, R. 1976. Language-bound and lan­ chol.-General 107:287-308
guage-optional processing. Presented at Gelman, R. 1978. Cognitive development.
Psychon. Soc., St. Louis Ann. Rev. Psychol 29:297-332
DeValois, R. L., Jacobs, F. H. 1968. Primate Glass, A. L., Holyoak, K. J. 1 975. Alterna­
color vision. Science 162:533-40 tive conceptions of semantic memory.
deVilIiers, J. 1 980. The process of rule learn­ Cognition 3 : 3 1 3-39
ing in child speech: A new look. In Chil­ Goldman, D., Homa, D. 1977. Integrative
and metric properties of abstracted in­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

dren 'sLanguage, ed. K. E. Nelson, Vol.

2. New York: Gardner. In press formation as a function of category dis­
Dougherty, J. W. D. 1978. Salience and rela­ criminability, instance variability, and
tivity in classification. Am. Ethnol. 5 : experience. J. Exp. Psychol.-Hum.
Learn. Mem. 3:375-85
by University of Sussex on 10/12/12. For personal use only.

66-80
Dreyfus, H. L. 1979. What Computers Can 't Hampton, J. A. 1979. Polymorphous con­
Do: A Critique ofArtificial Reason. New cepts in semantic memory. J. Verb.
York: Harper & Row. 354 pp. 2nd ed . Learn. Verb. Behav. 1 8 :441-61
Erickson, J. R., Jones, M. R. 1978. Thinking. Hayes-Roth, B., Hayes-Roth, F. 1977. Con­
cept learning and the recognition and
Ann. Rev. Psychol. 29:6 1-90
classification of exemplars. J. Verb.
Erreich, A., Valian, V. 1979. Children's inter­
Learn. Verb. Behav. 16:32 1-38
nal organization of locative categories.
Heider, E. R. 1972. Universals in color nam­
Child Dev. 50: 107 1-77
ing and memory. J. Exp. Psychol 93:
Fillmore, C. D. 1975. An alternative to
1 0-20
checklist theories of meaning. Proc. Jst
Hochberg, J., Galper, R. E. 1967. Recogni­
Ann. Meet. Berkeley Ling. Soc. 1 :
tion of faces I: An exploratory study.
123-31
Bull. Psychon. Soc. 9:619-20
Fillmore, C. D. 1977. Topics in lexical se­
Homa, D., Cross, J., Cornell, D., Goldman,
mantics. In Current Issues in Linguistic D., Shwartz, S. 1973. Prototype ab­
Theory, ed. R. W. Cole, pp. 76-138.
straction and classification of new in­
Wilmington, Ind: Indiana Univ. Press.
stances as a function of number of in­
303 pp.
stances defining the prototype. J. Exp.
Flavell, J. H. 1977. Cognitive Development. Psychol 101 : 1 1 6-22
Englewood Cliffs, NJ: Prentice Hall. Homa, D., Vosburgh, R. 1976. Category
286 pp. breadth and the abstraction of
Flavell, J. H., Wellman, H. M. 1976. Meta­ prototypical information. J. Exp. Psy­
memory. In Perspectives on the Develop­ chol-Hum. Learn. Mem. 2:322-30
ment ofMemory and Cognition, ed. R. I:Iull, C. L. 1 920. Quantitative aspects of the
V. K ail Jr., J. W. Hagen, pp. 3-33. evolution of concepts. Psychol. Monogr.
Hillsdale, NJ: Erlbaum. 498 pp. Vol. 28 (whole No. 123). 86 pp.
Fodor, J. 1970. Three reasons for not deriv­ Hunn, E. 1 975. Cognitive process�s in folk
ing 'kill' from 'cause to die'. Ling. Inq. ornithology: The identification of gulls.
1 :429-38 Work. pap. No. 42. Lang. Behav. Res.
Fodor, J. 1 9 8 1 . The current status of the in­ Lab., Univ. Calif., Berkeley
nateness controversy. In Representa­ Hunn, E. 1976. Toward a perceptual model
tions, ed. J. Fodor. Vermont: Bradford of folk biological classification. Am.
Books. In press Ethnol. 3:508-24
Fodor, J., Garrett, M., Walker, E., Parkes, R. Hupp, S. C., Mervis, C. B. 1 98 1 . Acquisition
1980. Against definitions. Cognition of Basic Object Categories by Severely
Vol. 8. In press Handicapped Children. Submitted for
Franks, J. J., Bransford, J. D. 1 97 1 . Abstrac­ publication.
tion of visual patterns. J. Exp. Psychol. Husaim, J. S., Cohen, L. B. 1980. Infant
90:65-74 learning of ill-defined categories. Pre­
Gamer, W. R. 1970. The stimulus in infor­ sented at Int. Conf. Infant Stud. New
mation processing. Am. PsychoL 25: Haven
3 50-58 Hutchinson, J. W., Lockhead, G. R. 1 977.
Gamer, W. R. 1974. The Processing ofInfor- Similarity as distance: A structural
 1 12 MERVIS & ROSCH

principle for semantic memory. J. Exp. Lakotf, G. 1977. On linguistic gestalts.

Psychol.-Hum. Learn. Mem. 3:660-78 Chicago Ling. Soc. Pap. 1 3:236-87
Hyman, R., Frost, N. H. 1975. Gradients and Leach, E. 1 964. Anthropological aspects of
schema in pattern recognition. In Atten­ language; Animal categories and verbal
tion and Performance, ed. P. M. A. abuse. In New Directions in ihe Study of
Rabbit, S. Dornig, 5:630-54. New Language, ed. E. H. Lenneberg, pp. 23-
York: Academic. 732 pp. 63. Cambridge, Mass: MIT Press.
James, W. I 890a. The Principles of Psy­ 1 94 pp.
chology, Vol. I. New York: Dover. Lippmann, W. 1922. Public Opinion. New
689 pp. York: Harcourt, Brace. 427 pp.
James, W. 1 890b. The Principles of Psy­ Locke, J. 1 690. An Essay Concerning Human
chology, Vol. 2. New York: Dover. Understanding. Annotated by A. C.
688 pp. Fraser, 1959. New York: Dover.
Kahneman, .D., Tversky, A. 1973. On the 495 pp.
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

psychology of prediction. Psychol. Rev. Lockhead, G. R. 1972. Processing dimen­

80:237-5 1 . sional stimuli: A note. Psychol. Rev.
Katz, J. J. 1 972. Semantic Theory. New 79:410-- 1 9
York: Harper & Row. 464 pp. MacKenzie, D. L., Palermo, D . S . 1 98 1 . The
by University of Sussex on 10/12/12. For personal use only.

Kay, P., McDaniel, C. K. 1978. The linguis­ Effect of Context on Semantic Catego­
tic significance of the meanings of basic rization. Submitted for publication
color terms. Language 54:610-46 Maratsos, M. P., Chalkley, M. A. 1980. The
Keil, F. C. 1979. Semantic and Conceptual internal language of children's syntax:
Development: An Ontological Perspec­ The ontogenesis and representation of
syntactic categories. See deVilliers
tive. Cambridge, Mass: Harvard Univ.
1 980. In press
Press. 2 1 4 pp.
McCauley, c., Stitt, C. L. 1978. An individ­
Keller, D. , Kellas, G. 1978. Typicality as a
ual and quantitative measure of stereo­
dimension of encoding. J. Exp. Psy­
types. J. Pers. Soc. Psychol. 36:929-40
chol.-Hum. Learn. Mem. 4:78-85
McCauley, c., Stitt, C. L., Segal, M. 1 980.
Kemler, D. G., Smith, L. B. 1978. Is there a
Stereotyping: From prejudice to predic­
developmental trend from integrality to
tion. Psychol. Bull 87: 195-208
separability in perception? J. Exp. Child
McCawley, J. 1971. Prelexical syntax.
PsychoL 26:498-507
Georgetown Univ. Monogr. Ser. Lang.
Kemler, D. G., Smith, L. B. 1979. Accessing
Ling. 2 3 : 1 9-33
similarity and dimensional relations:
McCawley, J. 1978. Conversational implica­
Effects of integrality and separability on
ture and the lexicon. In Syntax and Se­
the discovery of complex concepts. J. mantics, ed. P. Cole, 9:245-59. New
Exp. Psychol.-General 108:1 33-50 York: Academic. 340 pp.
Kempton, W. 1978. Category grading and McCloskey, M. E., Glucksberg, S. 1978. Nat­
taxonomic relations: A mug is a sort of ural categories: Well defined or fuzzy
a cup. Am. Ethnol. 5:44-65 sets? Mem. Cognit. 6 :462-72
Kinchla, R. A., Wolf, J. M. 1979. The order McCloskey, M. E., Glucksberg, S. 1979. De­
of visual processing: "Top-down," "bot­ cision processes in verifying category
tom-up," or "middle-out." Percept. Psy­ membership statements: Implications
chophys. 25:225-31 for models of semantic memory. Cognit.
Kintsch, W. 1 9 80. Semantic memory: A Psychol. 1 1 : 1-37
tutorial. In Attention and Performance Medin, D. L., Schaffer, M. M. 1978. Context
VIIL ed. R. S. Nickerson. Hillsdale, theory of classification learning. Psy­
NJ: Erlbaum. In press chol. Rev. 85:207-38
Kogan, N. 1 97 1 . Educational implication of Mervis, C. B. 1980. Category structure and
cognitive styles. In Psychology and Edu­ the development of categorization. In
cational Practice, ed. G. Lesser, pp. Theoretical Issues in Reading Compre­
242-92. Glenview, Ill: Scott, Foresman. hension, ed. R. Spiro, B. C. Bruce, W.
580 pp. F. Brewer. Hillsdale, NJ: Erlbaum. In
Labov, W. 1973. The boundaries of words press
and their meanings. In New Ways of Mervis, C. B., Catlin, J., Rosch, E. 1975. De­
Analyzing Variation in English, ed. C. J. velopment of the structure of color cate­
Bailey, R. Shuy, pp. 340-73. Washing­ gories. Dev. PsychoL 1 1 :54-60
ton: Georgetown Univ. Press. 373 pp. Mervis, C. B., Catlin, J., Rosch, E. 1976. Re­
Lakotf, G. 1973. Hedges: A study in meaning lationships among goodness-of-exam­
criteria and the logic of fuzzy concepts. pie, category norms, and word fre­
J. Philos. Logic 2:458-508 quency. Bull Psychon. Soc. 7:283-84
 CATEGORIZAnON OF OBJECTS 1 13

Mervis, C. B., Crisafi, M. A. 1981. The Per­ ration in Cognition, ed. D. A. Norman,
ceptual-Cognitive Primacy of Basic D. E. Rumelhart, LNR Research
Level Categories. In preparation Group, pp. 279-307. San Francisco:
Mervis, C. B., Mirman, M. C. 1 9 8 1 . The Freeman. 430 pp.
Effects of Category Size and Distortion Palmer, S. E. 1978. Fundamental aspects of
Level on Concept Abstraction. In prepa­ cognitive representation. See Brooks
ration 1978, pp. 259-303
Mervis, C. B., Pani, J. R. 1 980. Acquisition Piagei:, J. 1 970. Piaget's theory. In Carmicha­
of basic object categories. Cognit. Psy­ el's Manual of Child Psychology, ed. P.
chol. Vol. 12. In press H. Mussen, 1 :703-32. New York: Wi­
Meyer, D. E. 1970. On the representation and ley. 1 5 19 pp. Piaget, J. 1972. The Princi­
retrieval of stored semantic infonna­ pies of Genetic Epistemology. New
tion. Cognit. PsychoL 1 :242-300 York: Basic Books. 98 pp.
Miller, G. A., Johnson-Laird, P. N. 1976. Pittenger, J. B., Shaw, R. E. 1975. Aging
Language and Perception. Cambridge, faces as visual-elastic events: Implica­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

Mass: Harvard .Univ. Press. 760 pp. tions for a theory of nonrigid shape per­
Mirman, M. C. 1978. The Effects ofNumber ception. J. Exp. Psychol. -Hum. Percept.
and Distortion Level of Category Exem­ Perform. 1 :374-82
plars on Concept Abstraction. Honors Pittenger, J. B., Shaw, R. E., Mark, L. S.
by University of Sussex on 10/12/12. For personal use only.

BS thesis. Univ. Illinois, Urbana-Cham­ 1979. Perceptual information for the

paign age level of faces as a higher order in­
Murphy, G., Smith, E. E. 1 98 1 . Levels of Cat­ variant of growth. J. Exp. Psychol. ­
egorization and Object Identification. In Hum. Percept. Perform. 5:478-93
preparation Posnansky, C. J. 1973. Category norms for
Navon, D. 1977. Forest before trees: The verbal items in 25 categoriesfor children
precedence of global features in visual in grades 2-6. Prog. Cognit. Factors
perception. Cognit. PsychoL 9:353-83 Hum. Learn. Mem., Rep. 7, Univ.
Neimark, E. D. 1974. Natural language con­ Colo., Boulder
cepts: Additional evidence. Child Dev. Posnansky, C. J., Neumann, P. G. 1976. The
45:508- 1 1 abstraction of visual prototypes by chil­
Neimark, E. D., Santa, J . L. 1975. Thinking dren. J. Exp. Child Psychol. 2 1 :367-79
and concept attainment. Ann. Rev. Psy­ Posner, M. I. 1969. Abstraction and the pro­
chol. 26: 1 73-205 cess of recognition. In The Psychology of
Nelson, K. 1973. Some evidence for the cog­ Learning and Motivation, ed. G. H.
nitive primacy of categorization and its Bower, J. T. Spence, 3:44-96. New
functional basis. Merrill-Palmer Q. 19: York: Academic. 398 pp.
2 H9 Posner, M. I., Goldsmith, R., Welton, K. E.,
Nelson, K. 1974. Variations in children's Jr. 1967. Perceived distance and the
concepts by age and category. Child classification of distorted patterns. J.
Dev. 45:577-84 Exp. Psychol. 73:28-38
Nelson, K. E., Nelson, K. 1978. Cognitive Posner, M. I., Keele, S. W. 1968. On the
pendulums and their linguistic realiza­ genesis of abstract ideas. J. Exp. Psy­
tion. In Children 's Language, ed. K. E. choL 77:353-63
Nelson, 1 :223-85. New York: Gardner. Potter, M. C, Faulconer, B. A. 1979. Under­
567 pp. standing noun phrases. J. Verb. Learn.
Neumann, P. G. 1974. An attribute fre­ Verb. Behav. 1 8:509-2 1
quency model for the abstraction of Rabinow, P., Sullivan, W. M., eds. 1979. In­
prototypes. Mem. Cognit. 2:241-48 terpretive Social Science: A Reader. Ber"
Neumann, P. G. 1977. Visual prototype for­ keley: Univ. Calif. Press. 367 pp.
mation with discontinuous representa­ Reber, A. S. 1976. Implicit learning of syn­
tion of dimensions of variability. Mem. thetic llinguages: The role of instruc­
Cognit. 5 : 1 87-97 tional set. J. Exp. Psychol.-Hum.
Newport; E. L., Bellugi, U. 1978. Linguistic
. Learn. Mem. 2:88-94
expression of category levels in a visual­ Reed, S. K. 1972: Pattern recognition and '
gestural language: A Bower is a Bower is categorization.· Cognit. PsychoL 3:382-
a flower. See Brooks 1978, pp. 49....7 1 407
Oden, G. C., Massaro, D . W. 1978. Integra­ Reed, S. K. 1 973. Psychological Processes in
tion of featural iilformation in speech Pattern Recognition. New York: Aca­
perception. PsychoL Rev. 85:172-91 demic. 244 pp.
Palmer, S. E. 1975. Visual perception and Reed, S. K., Fnedman, M. P. 1973. Per­
world knowledge: Notes on a model of ceptual vs. conceptual categorization.
sensory-cognitive interaction. In Explo- Mem. Cognit. 1 : 1 57-63
 1 14 MERVIS & ROSCH

Ricciuti, H. N. 1965. Object grouping and 2 year olds. Presented at Bien Meet.
selective ordering behavior in infants 12 Soc. Res. Child Dev., New Orleans
to 24 months old. Merrill-Palmer Q. Ross, J. 1972. The category squish: Endsta­
1 1 : 1 29-48 tion Hauptwort. Chicago Ling. Soc.
Riley, D. A., Lamb, M. R. 1979. Stimulus Pap. 8:3 1 6-29
generalization. In Perception and Its De­ Saltz, E., Dixon, D., Klein, S., Becker, G.
velopment: A Tribute to Eleanor J. Gib­ 1 977. Studies of natural language con­
son, ed. A. D. Pick, pp. 7-37. Hillsdale, cepts: III. Concept overdiscrimination
NJ: Erlbaum. 258 pp. in comprehension between two and four
Rips, L. J. 1975. Inductive judgments about years of age. Child Dev. 48:1 682-85
natural categories. J. Verb. Learn. Verb. Saltz, E., Soller, E., Sigel, I. E. 1972. The
Behav. 14:665-8 1 development of natural language con­
Rips, L. J., Shoben, E. J., Smith, E. E. 1973. cepts. Child Dev. 43: 1 19 1 - 1 202
Semantic distance and the verification Schank, R. C. 1973. Identification of concep­
tualizations underlying natural lan­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

of semantic relations. J. Verb. Learn.

Verb. Behav. 12: 1-20 guage. In Computer Models of Thought
Rock, I. 1973. Orientation and Form. New and Language, ed. R. C. Schank, K. M.
York: Academic Press. 165 pp. Colby, pp. 1 87-247. San Francisco:
Freeman. 454 pp.
Rosch, E. H. 1973a. Natural categories. Cog­
by University of Sussex on 10/12/12. For personal use only.

nit. PsychoL 4:328-50 Schank, R. C., Abelson, R. P. 1977. Scripts,

Plans, Goals and Understanding: An In­
Rosch, E. 1973b. On the internal structure of
quiry into Human Knowledge Struc­
perceptual and semantic categories. In
tures. Hillsdale, NJ:Erlbaum. 248 pp.
Cognitive Development and the Acquisi­
Shepard, R. N., Chipman, S. 1970. Second­
tion of Language, ed. T. E. Moore, pp.
order isomorphism of internal represen­
1 1 1-44. New York: Academic. 308 pp.
tations: Shapes of states. Cognit. Psy­
Rosch, E. 1974. Linguistic relativity. In Hu­
chol 1 : 1- 1 7
man Communication: Theoretical Ex­
Shepp, B. E . 1978. From perceived similarity
plorations, ed. A. Silverstein, pp. 95-
to dimensional structure: A new hy­
1 2 1 . Hillsdale, NJ: Erlbaum. 264 pp.
pothesis about perceptual development.
Rosch, E. 1975a. Cognitive reference points.
See Brooks 1978, pp. 135-67
Cognit. Psychol 7:532-47
Shepp, B. E., Swartz, K. B. 1976. Selective
Rosch, E. 1975b. Cognitive representations attention and the processing of integral
of semantic categories. J. Exp. Psychol.­ and nonintegral dimensions: A develop­
General 104:1 92-233 mental study. J. Exp. Child Psychol.
Rosch, E. 1975c. The nature of mental codes 22:73-85
for color categories. J. Exp. Psychol.­ Simpson, G. G. 1961. Principles of Animal
Hum. Percept. Perform. 1 :303-22 Taxonomy. New York: Columbia Univ.
Rosch, E. 1977. Human categorization. In Press. 247 pp.
Studies in Cross-Cultural Psychology, Smith, E. E. 1978. Theories of semantic
ed. N. Warren, 1 :3-49. London: Aca­ memory. In Handbook ofLearning and
demic. 2 1 2 pp. Cognitive Processes, ed. W. K. Estes,
Rosch, E. 1978. Principles of categorization. 6: 1-56. Hillsdale, NJ: Erlbaum. 3 1 1 pp.
See Brooks 1978, pp. 27-48 Smith, E. E., Balzano, G. 1977. Concepts and
Rosch, E. 1979. Attributes. Presented at prototypes. Colloquium, Harvard Univ.
Conf. Word & Object, Stanford, Calif. Smith, E. E., Medin, D. L. 198 1 . The Psy­
Rosch, E., Mervis, C. B. 1975. Family resem­ chology of Conceptual Processes. Cam­
blances: Studies in the internal structure bridge, Mass: Harvard Univ. Press. In
of categories. Cognit. Psychol. 7:573- press
605 Smith, E. E., Rips, L. J., Shoben, E. J. 1974.
Rosch, E., Mervis, C. B., Gray, W. D., John­ Semantic memory and psychological se­
son, D. M., Boyes-Braem, P. 1 976a. Ba­ mantics. In The Psychology ofLearning
sic objects in natural categories. Cognit. and Motivation, ed. G. H. Bower, 8 : 1 -
PsychoL 8:382-439 45. New York: Academic. 305 pp.
Rosch, E., Simpson, C., Miller, R. S. 1 976b. Smith, L. B., Kemler, D. G. 1977. Develop­
Structural bases of typicality effects. J. mental trends in free classification: Evi­
Exp. Psychol-Hum. Percept. Perform. dence for a new conceptualization of
2:491-502 perceptual development. J. Exp. Child
Rosner, S. R., Hayes, D. S. 1977. A develop­ Psychol. 24:279-98
mental study of category item produc­ Smith, L. B., Kemler, D. G. 1978. Levels of
tion. Child Dev. 48: 1 062-65 experienced dimensionality in children
Ross, G. 1977. Concept categorization in 1 to and adults. Cognit. Psychol. 10:502-32
 CATEGORIZATION OF OBJECTS 1 15

Sneath, P. H. A., Sokal, R. R. 1973. Numeri­ Behav. 1 8 : 109-27

cal Taxonomy. San Francisco: Free­ Williams, T. M., Fryer, M. 1., Aiken, 1. S.
man. 573 pp. 1977. Development of visual pattern
Sokal, R. R. 1974. Classification: Purposes, classification in preschool children:
principles, progress, prospects. Science Prototypes and distinctive features.
185: 1 1 1 5-23 Dev. Psychol. 1 3:577-84
Strauss, M. S. 1979. Abstraction ofprototypi­ Winograd, T. 1972. Understanding natural
cal information by adults and 10- language. Cognit. PsychoL 3 : 1 - 1 9 1
month-old infants. J. Exp. PsychoL­ Winograd. T. 1975. Frame representations
Hum. Learn. Mem. 5 : 6 1 8-32 and the declarative-procedural contro­
Stross, B. 1973. Acquisition of botanical ter­ versy. In Representation and Under­
minology by Tzeltal children. In Mean­ standing; Studies in Cognitive Science,
ing in Mayan Languages. ed. M. Ed­ ed. D. G. Bobrow, A. Collins, pp. 1 85-
monson, pp. 107-42. The Hague: Mou­ 2 1 0. New York: Academic. 427 pp.
ton. 256 pp. Winston, P. H. 1975. Learning structural de­
Annu. Rev. Psychol. 1981.32:89-115. Downloaded from www.annualreviews.org

Tversky, A. 1 977. Features of similarity. Psy­ scription from examples. In The Psy­
chol. Rev. 84:327-52 chology of Computer Vision. ed. P. H .
Tversky, A., Gati, I. 1978. Studies of simi­ Winston, pp. 1 57-2 10. New York:
larity. See Brooks 1978, pp. 79-98 McGraw-Hill. 282 pp.
by University of Sussex on 10/12/12. For personal use only.

Wanner, E. 1979. Discussion of prototypes. Witkowski, S. R., Brown, C. H. 1978. Lex­

Presented at Meet. Soc. Philos. Psy­ ical universals. Ann. Rev. Anthropol.
chol.. New York 7:427-5 1
Whitfield, T. W. A., Slatter, P. E. 1979. The Wittgenstein, L. 1953. Philosophical Investi­
effects of categorization and prototypi­ gations. New York: Macmillan. 232 pp.
cality on aesthetic choice in a furniture Yin, R. K. 1969. Looking at upside-down
selection task. Br. J. PsychoL 70:65-75 faces. J. Exp. Psychol. 8 1 : 14 1-45
Whitten, W. B. II, Suter, W. N., Frank, M. Zadeh, 1. A. 1965. Fuzzy sets. Inf. Control
1. 1979. Bidirectional synonym ratings 8:338-53
of 464 noun pairs. 1. Verb. Learn. Verb.