Principles
of Categorization
Eleanor Rosch, 1978
University of
California, Berkeley
From Readings in
Cognitive Science, a Perspective from Psychology and Artificial Intelligence,
Allan Collins &
Edward E. Smith, Morgan Kaufmann Publishers, San Mateo, California, 1988, pp
312-322
The following is a taxonomy of the animal kingdom. It has
been attributed to an ancient Chinese encyclopedia entitled the Celestial
Emporium of Benevolent Knowledge:
On those remote pages it is written
that animals are divided into (a) those that belong to the Emperor, (b)
embalmed ones, (c) those that are trained, (d) suckling pigs, (e) mermaids, (f)
fabulous ones, (g) stray dogs, (h) those that are included in this
classification, (i) those that tremble as if they were mad, j) innumerable
ones, (k) those drawn with a very fine camel's hair brush, (l) others, (m)
those that have just broken a flower vase, (n) those that resemble flies from a
distance (Borges, 1966, p. 108).
Conceptually, the most interesting aspect of this classification
system is that it does not exist. Certain types of categorizations may appear
in the imagination of poets, but they are never found in the practical or
linguistic classes of organisms or of man-made objects used by any of the
cultures of the world. For some years, I have argued that human categorization
should not be considered the arbitrary product of historical accident or of
whim but rather the result of psychological principles of categorization, which
are subject to investigation. This chapter is a summary and discussion of those
principles.
The chapter is divided into five parts.
The first part presents the two general principles that are
proposed to underlie categorization systems.
The second part shows the way in which these principles
appear to result in a basic and primary level of categorization in the levels
of abstraction in a taxonomy: It is essentially a summary of the research
already reported on basic level objects (Rosch et al., 1976). Thus the second
section may be omitted by the reader already sufficiently familiar with that
material.
The third part relates the principles of categorization to
the formation of prototypes in those categories that are at the same level of
abstraction in a taxonomy. In particular, this section attempts to clarify the
operational concept of prototypicality and to separate that concept from claims
concerning the role of prototypes in cognitive processing, representation, and
learning for which there is little evidence.
The fourth part presents two issues that are problematical
for the abstract principles of categorization stated in Part 1: (1) the
relation of context to basic level objects and prototypes; and (2) assumptions
about the nature of the attributes of real-world objects that underlie the
claim that there is structure in the world.
The fifth part is a report of initial attempts to base an
analysis of the attributes, functions, and contexts of objects on a
consideration of objects as props in culturally defined events.
It should be noted that the issues in categorization with
which we are primarily concerned have to do with explaining the categories
found in a culture and coded by the language of that culture at a particular
point in time. en we speak of the formation of categories, we mean their
formation in the culture. This point is often misunderstood. The principles of
categorization proposed are not as such intended to constitute a theory of the
development of categories in children born into a culture nor to constitute a
model of how categories are processed (how categorizations are made) in the
minds of adult speakers of a language.
THE PRINCIPLES
Two general and basic principles are proposed for the
formation of categories:
The first has to do with the function of category systems
and asserts that the task of category systems is to provide maximum information
with the least cognitive effort.
The second principle has to do with the structure of the
information so provided and asserts that the perceived world comes as structured
information rather than as arbitrary or unpredictable attributes. Thus maximum
information with least cognitive effort is achieved if categories map the
perceived world structure as closely as possible. This condition can be
achieved either by the mapping of categories to given attribute structures or
by the definition or redefinition of attributes to render a given set of
categories appropriately structured.
These principles are elaborated in the following.
Cognitive Economy.
The first principle contains the almost common-sense notion
that, as an organism, what one wishes to gain from one's categories is a great
deal of information about the environment while conserving finite resources as
much as possible. To categorize a stimulus means to consider it, for purposes
Purposes of that categorization, not only equivalent to
other stimuli in the same category but, also different from stimuli not in that
category, On the one hand, it would appear to the organism's advantage to have
as many properties as possible predictable from Knowing any one property, a
principle that would lead to formation of large numbers of categories with as
fine discriminations between categories as possible. On the other hand , one purpose of categorization is to reduce the
infinite differences among stimuli to behaviorally and cognitively usable
proportions. It is to the organism's advantage not to differentiate one
stimulus from others when that differentiation is irrelevant to the purposes at
hand.
Perceived World Structure.
The second principle of categorization asserts that unlike
the sets of stimuli used in traditional laboratory concept attainment tasks,
the perceived world is not an unstructured total set of equiprobable co-occurring
attributes Rather, the material objects of the world are perceive to possess
(in Garner's, 1974, sense) high correlational structure. That is, given a
knower who perceives the complex attributes of feathers fur and wings, it is an
empirical fact provided by the perceived world that wings co-occur with
feathers more than with fur. And given an actor with the motor programs for
sitting, it is a fact of the perceived world that objects with the perceptual
attributes of chairs are more likely to have functional sit-on-able-ness than
objects with the appearance of cats. In short, combinations of what we perceive
as the attributes of real objects do not occur uniformly. Some pairs, triples,
etc., are quite probable, appearing in combination sometimes with one,
sometimes another attribute; others are rare; others logically cannot or
empirically do not occur.
It should be emphasized that we are talking about the
perceived world and not a metaphysical world without a knower. What kinds of
attributes can be perceived are, of course, species-specific. A dog's sense of smell is more highly
differentiated than a human's, and the structure of the world for a dog must
surely include attributes of smell that we, as a species, are incapable of
perceiving. Furthermore, because a dog's body is constructed differently from a
human's, its motor interactions with objects are necessarily differently
structured. The "out there" of a bat, a frog, or a bee is surely more
different still from that of a human. What attributes will be perceived given the
ability to perceive them is undoubtedly determined by many factors having to do
with the functional needs of the knower interacting with the physical and
social environment. One influence on how attributes will be defined by humans
is clearly the category system already existent in the culture at a given time.
Thus, our segmentation of a bird's body such that there is an attribute called
"wings" may be influenced not only by perceptual factors such as the
gestalt laws of form that would lead us to consider the wings as a separate
part (Palmer, in press) but also by the fact that at present we already have a
cultural and linguistic category called "birds." Viewing attributes
as, at least in part, constructs of the perceiver does not negate the
higher-order structural fact about attributes at issue, namely that the
attributes of wings and that of feathers do co-occur in the perceived world.
These two basic principles of categorization, a drive toward
cognitive economy combined with structure in the perceived world, have
implications both for the level of abstraction of categories formed in a
culture and for the, internal structure of those categories once formed.
For purposes of explication, we may conceive of category
systems as having both a verbal and horizontal dimension. The vertical
dimension concerns the level of inclusiveness of the category - the dimension
along which the terms collie, dog, mammal, animal, and living thing vary. The
horizontal dimension concerns the segmentation of categories at the same level
of inclusiveness - the dimension on which dog, cat car, bus chair, and sofa
vary. The implication of the two principles of categorization for the vertical
dimension is that not all possible levels of categorization are equally good or
useful; rather, the most basic level of categorization will be the most
inclusive (abstract) level at which the categories can mirror the structure of
attributes perceived in the world. The implication of the principles of
categorization for the horizontal dimension is that to increase the
distinctiveness and flexibility of categories, categories tend to become
defined in terms of prototypes or prototypical instances that contain the
attributes most representative of items inside and least representative of
items outside the category.
THE
VERTICAL DIMENSION OF CATEGORIES: BASIC-LEVEL OBJECTS
In a programmatic series of experiments, we have attempted
to argue that categories within taxonomies of concrete objects are structured
such that there is generally one level of abstraction at which the most basic
category cuts can be made (Rosch et al., 1976a). By category is meant a number
objects that are considered equivalent. Categories are generally designated by
names (e.g., dog, animal) A taxonomy is
a system by which categories are related to one another by means of class
inclusion, The greater the inclusiveness of a category within a taxonomy, the
higher the level of abstraction., Each category within a taxonomy is entirely
included within one other category (unless it is the highest level category)
but is not exhaustive of that more inclusive category (see Kay, 1971). Thus the term level of abstraction
within a taxonomy refers to a particular level of inclusiveness. A familiar
taxonomy is the Linnean system for the classification of animals.
Our claims concerning a basic level of abstraction can be
formalized in terms of cue validity (Rosch et al., 1976a) or in terms of the
set theoretic representation of similarity provided by Tversky (1977, and
Chapter 4 in this volume). Cue validity is a probabilistic concept; the
validity of a given cue x as a predictor of a given category y (the conditional
probability of y/x) increases as the frequency with which cue x is associated
with category y increases and decreases as the frequency with which cue x is
associated with categories other than y increases (Beach, 1964a, 1964b; Reed,
1972). The cue validity of an entire category may be defined as the summation
of the cue validities for that category of each of the attributes of the category.
A category with high cue validity is, by definition, more differentiated from
other categories than one of lower cue validity. The elegant formulization that
Tversky provides in Chapter 4 is in terms of the variable "category
resemblance," which is defined as the weighted sum of the measures of all
of the common features within a category minus the sum of .the measures of all
of the distinctive features., Distinctive features include those that belong to
only some members of a given category as well as those belonging to contrasting
categories. Thus Tversky's formalization does not weight the effect of contrast
categories as much as does the cue validity formulation. Tversky suggests that
two disjoint classes tend to be combined whenever the weight of the added
common features exceeds the weight of the distinctive features.
A working assumption of the research on basic objects that
(1) in the perceived world, information-rich bundles of perceptual and
functional attributes occur that form natural discontinuities, and that
(2),basic cuts in categorization are made at these discontinuities. Suppose
that basic objects, (e.g., chair, car), are at the most inclusive level at
which there are attributes common to all or most members of the category. Then
both total cue validities and category resemblance are maximized at that level
of abstraction at which basic objects are categorized. This is, categories one
level more abstract will be superordinate categories (e.g., furniture, vehicle)
whose members share only a few attributes among each other. Categories below
the basic level will be bundles of common and, thus, predictable attributes and
functions but contain many attributes that overlap with other categories (for
example, kitchen chair shares most of its attributes with other kinds of
chairs).
Superordinate categories have lower total cue validity and
lower category resemblance than do basic-level categories, because they have
fewer common attributes; in fact, the category resemblance measure of items
within the superordinate can even be negative due to the high ratio of
distinctive to common features. Subordinate categories have lower total cue
validity than do basic categories, because they also share most attributes with
contrasting subordinate categories; in Tversky's terms, they tend to be
combined because the weight of the added common features tend to exceed the
weight of the distinctive features, That basic objects are categories at the
level of abstraction that maximizes cue validity and maximizes category
resemblance is another way of asserting that basic objects are the categories
that best mirror the correlational structure of the environment.
We chose to look at concrete objects because they appeared
to be a domain that was at once an indisputable aspect of complex natural
language classifications yet at the same time were amenable to methods of
empirical analysis. In our investigations of basic categories, the
correlational structure of concrete objects was considered to consist of a
number of inseparable aspects of form and
function, any one of which could serve as the starting point
for analysis. Four investigations provided converging operational definitions
of the basic level of abstraction: attributes in common, motor movements in
common, objective similarity in shape, and identifiability of averaged shapes.
Common Attributes.
Ethnobiologists had suggested on the basis of linguistic
criteria and field observation that the fobs was the level of classification at
which organisms had bundles of attributes in common and maximum discontinuity
between classes (see Chapter 1). The purpose of our research was to provide a
systematic empirical study of the co-occurrence of attributes in the most
common taxonomies of biological and man-made objects in our own culture.
The hypothesis that basic level objects are the most
inclusive level of classification at which objects have numbers of attributes
in common was tested for categories at three levels of abstraction for nine
taxonomies: tree, fish, fruit, musical instruments, tool, clothing furniture
and vehicle. Examples of the three levels for one biological and one
nonbiological taxonomy are shown in Table 2.1. Criteria for choice of these
specific items were that the taxonomies contain the most common (defined by
word frequency) categories of concrete nouns in English, that the levels of
abstraction bear simple class-inclusion relations to each other, and that
those class-inclusion relations be generally known to our subjects (be agreed
upon by a sample of native English speakers) The middle level of abstraction
was the hypothesized basic level: For nonbiological taxonomies, this
corresponded to the intuition of the experimenters (which also turned out to be
consistent with Berlin's linguistic criteria); for biological categories, we
assumed that the basic level would be the level of the folk generic.
Subjects received sets of words taken from these nine
taxonomies; the subject's task was to list all of the attributes he could think
of that were true of the items included in the class of things designated by
each object name. Thus, for purposes of this study, attributes were defined
operationally as whatever subjects agreed them to be with no implications for
whether such analysis of an object could or could not be perceptually
considered prior to knowledge of the object itself. Results of the study were
as predicted: Very few attributes were listed for the superordinate categories,
a significantly greater number listed for the supposed basic-level objects, and
not significantly more attributes listed for subordinate level objects than
for basic-level. An additional study showed essentially the same attributes
listed for visually present objects as for the object names. The angle
unpredicted result was that for the three biological taxonomies, the basic
level, as defined by numbers of attributes in common, did not occur at the
level of the folk generic but appeared at the level we had originally expected
to be superordinate (e.g., tree rather than oak).
TABLE 2.1 Examples of Taxonomies Used in Basic Object
Research
|
Superordinate
|
Basic Level
|
Subordinate
|
|
Furniture
|
Chair
|
Kitchen chair
|
|
|
|
Living-room chair
|
|
|
Table
|
Kitchen table
|
|
|
|
Dining-room table
|
|
|
Lamp
|
Floor lamp
|
|
|
|
Desk lamp
|
|
Tree
|
Oak
|
White oak
|
|
|
|
Red oak
|
|
|
Maple
|
Silver maple
|
|
|
|
Sugar maple
|
|
|
Birch
|
River birch
|
|
|
|
White birch
|
Motor Movements.
Inseparable from the perceived attributes of objects are the
ways in which humans habitually use or interact with those objects. For
concrete objects, such interactions take the form of motor movements. For
example, when performing the action of sitting down on a chair, a sequence of
body and muscle movements are typically made that are inseparable from the
nature of the attributes of chairs - legs, seat, back, etc. This aspect of
objects is particularly important in light of the role that sensory-motor
interaction with the world appears to play in the development of thought
(Brunei, Olver, & Greenfield, I966; Nelson, I974; Piaget, I952).
In our study of motor movements, each of the sets of words
used in the previous experiment was administered to new subjects. A subject
was asked to describe, in as much finely analyzed detail as possible, the
sequences of motor movements he made when using or interacting with the object.
Tallies of agreed upon listings of the same movements of the same body part in
the same part of the movement sequence formed the unit of analysis. Results
were identical to those of the attribute listings; basic objects were the most
general classes to have motor sequences in common. For example, there are few
motor programs we irry out to items of furniture in general and several
specific motor programs +cried out in regard to sitting down on chairs, but we
sit on kitchen and livingroom chairs using essentially the same motor
programs.
Similarity in Shapes.
Another aspect of the meaning of a class of objects is the
appearance of the objects in the class. In order to be able to analyze
correlational structures by different but converging methods, it was necessary
to find a method of analyzing similarity in the visual aspects of the objects
that was not dependent on subjects' descriptions, that was free from effects of
the object's name (which would not have been the case for subjects' ratings of
similarity), and that went beyond similarity of _analyzable, listable
attributes that had already been used in the first study described. For this
purpose, outlines of the shape of two-dimensional representations of objects
were used, an integral aspect of natural forms. Similarity in shape was
measured by the amount of overlap of the two outlines when the outlines
(normalized for size and orientation) were juxtaposed.
Results showed that the ratio of overlapped to nonoverlapped
area when two objects from the same basic-level category (e.g., two cars) were
superimposed was far greater than when two objects from the same superordinate
category
were superimposed (e-g., a car and a motorcycle). Although
some gain in ratio of overlap to nonoverlap also occurred for subordinate
category objects (e.g., two sports cars), the gain obtained by shifting from
basic-level to subordinate objects was significantly less than the gain
obtained by shifting from superordinate to basic-level objects.
Identifiability of Averaged Shapes.
If the basic level is the most inclusive level at which
shapes of objects of a class are similar, a possible result of such similarity
may be that the basic level is also the most inclusive level at which an
averaged shape of an object can be recognized. To test this hypothesis, the
same normalized superimposed shapes used in the previous experiment were used
to draw an average outline of the overlapped figures. Subjects were then asked
to identify both the superordinate category and the specific object depicted.
Results showed that basic objects were the most general and inclusive
categories at which the objects depicted could be identified. Furthermore,
overlaps of subordinate objects were no more identifiable than objects at the
basic level.
In summary, our four converging operational definitions of
basic objects all indicated the same level of abstraction to be basic in our
taxonomies. Admittedly, the basic level for biological objects was not that
predicted by the folk genus; however, this fact appeared to be simply accounted
for by our subjects' lack of knowledge of the additional depth of real-world
attribute structure available at the level of the folk generic (see Rosch et
al., 1976a).
Implications for Other Fields
The foregoing theory of categorization and basic objects has
implications for several traditional areas of study in psychology; some of
these have been tested.
Imagery.
The fact that basic-level objects were the most inclusive
categories which an averaged member of the category could be identified
suggested at basic objects might be the most inclusive categories for which it
was possible form a mental image isomorphic to the appearance of members of the
class as whole. Experiments using a signal-detection paradigm and a priming
paradigm, both of which have been previously argued to be measures of imagery
(Peterson Graham, I974; Rosch, 1975c), verified that, in so far as it was
meaningful to e the term imagery, basic objects appeared to-be the most
abstract categories for which an image could be reasonably representative of
the class as a whole.
Perception.
From all that has been said of the nature of basic
classifications, would hardly be reasonable to suppose that in perception of
the world, objects were first categorized either at the most abstract or at the
most concrete level possible Two separate studies of picture verification
(Rosch et al., 1976a; Smith, Balzano, & Walker, I978) indicate that, in
fact, objects may be first :en or recognized as members of their basic category,
and that only with the ad of additional processing can they be identified as
members of their superordinate or subordinate category.
Development.
We have argued that classification into categories at the
basic level is overdetermined because perception, motor movements, functions,
and iconic images would all lead to the same level of categorization. Thus
basic objects should be the first categorizations of concrete objects made by
children. In fact, for our nine taxonomies, the basic level was the first
named. And even when naming was controlled, pictures of several basic-level
objects were sorted into groups "because they were the same type of
thing" long before such a technique of sorting has become general in
children.
Language.
From all that has been said, we would expect the most useful
and, thus, most used name for an item to be the basic-level name. In fact, we
found that adults almost invariably named pictures of the subordinate items of
the nine taxonomies at the basic level, although they knew the correct superordinate
and subordinate names for the objects. On a more speculative level, in the
evolution of languages, one would expect names to evolve first for basic level
objects, spreading both upward and downward as taxonomies increased in depth.
Of great relevance for this hypothesis are Berlin's (1972) claims for such a
pattern for the evolution of plant names, and our own (Rosch et al., 1976a) and
Newport and Bellugi's (Chapter 3, this volume) finding for American Sign
Language of the Deaf, that it was the basic-level categories that were most
often coded by single signs and super- and subordinate categories that were
likely to be missing. Thus a wide range of converging operations verify as
basic the same levels of abstraction.
THE HORIZONTAL DIMENSION:
INTERNAL STRUCTURE OF CATEGORIES: PROTOTYPES
Most, if not all, categories do not have clear-cut
boundaries. To argue that basic object categories follow clusters of perceived
attributes is not to say that such attribute clusters are necessarily
discontinuous.
In terms of the principles of categorization proposed
earlier, cognitive economy dictates that categories tend to be viewed as being
as separate from each other and as clear-cut as possible. One way to achieve
this is by means of formal, necessary and sufficient criteria for category
membership. The attempt to impose such-criteria on categories marks virtual
definitions in the tradition of Western reason. The psychological treatment of
categories in the standard concept-identification paradigm lies within this
tradition. Another way to achieve separateness and clarity of actually
continuous categories is by conceiving of each category in terms of its clear
cases rather than its boundaries. As Wittgenstein (1953) has pointed out, categorical
judgments become a problem only if one is concerned with boundaries - in the
normal course of life, two neighbors know on whose property they are standing
without exact demarcation of the boundary line. Categories can be viewed in
terms of their clear cases if the perceiver places emphasis on the
correlational structure of perceived attributes such that the categories are
represented by their most structured portions.
By prototypes of categories we have generally meant the
clearest cases of pry membership defined operationally by people's judgments of
goodness of membership in the category. A great deal of confusion in the
discussion of prototypes has arisen from two sources. First, the notion of
prototypes has tended to become reified as though it meant a specific category
member or mental structure. Questions are then asked in an either-or fashion
about whether something is or is not the prototype or part of the prototype in
exactly the same way in which the question would previously have been asked
about the category boundary. Such thinking precisely violates the
Wittgensteinian insight that we can judge how clear a case something is and
deal with categories on the basis of clear cases in the total absence of
information about boundaries. Second, the empirical findings about
prototypicality have been confused with theories of processing - that is, there
has been a failure to distinguish the structure of categories from theories
concerning the use of that structure in processing. Therefore, let us first
attempt to look at prototypes in as purely structural a fashion as possible. We
will focus on what may be said about prototypes based on operational
definitions and empirical findings alone without the addition of processing
assumptions.
Perception of typicality differences is, in the first place,
an empirical fact of people's judgments about category membership. It is by now
a well-documented finding that subjects overwhelmingly agree in their judgments
of how good an example or clear a case members are of a category, even for
categories about whose boundaries they disagree (Rosch, 1974, 1975b). Such
judgments are reliable even under changes of instructions and items (Rips,
Shoben, & Smith, 1973; Rosch, 1975b, 1975c; Rosch & Mervis, 1975). Were
such agreement and reliability in judgment not to have been obtained, there
would be no further point in discussion or investigation of the issue. However,
given the empirical verification of degree of prototypicality, we can proceed
to ask what principles determine which items will be judged the more
prototypical and what other variables might be affected by prototypicality.
In terms of the basic principles of category formation, the
formation of category prototypes should, like basic levels of abstraction, be
determinate and be closely related to the initial formation of categories. For
categories of concrete objects (which do not have a physiological basis, as
categories such as colors and forms apparently do - Rosch, 1974), a reasonable
hypothesis is that prototypes develop through the same principles such as
maximization of cue validity and maximization of category resemblance as those
principles governing the formation of the categories themselves.
In support of such a hypothesis, Rosch and Mervis (1975)
have shown that the more prototypical of a category a member is rated, the more
attributes it .has in common with other members of the category and the fewer
attributes in common with members of the contrasting categories. This finding
was demonstrated for natural language superordinate categories, for natural
language basic level categories, and for artificial categories in which the
definition of attributes and the amount of experience with items was completely
specified and controlled. The same basic principles can be represented in ways
other than through attributes in common. Because the present theory is a
structural theory, one aspect of it is that centrality shares the mathematical
notions inherent in measures like the mean and mode. Prototypical category
members have been found to represent the means of attributes that have a
metric, such as size (Reed, 1972; Rosch, Simpson, & Miller, 1976).
In short, prototypes appear to be just those members of a
category that most reflect the redundancy structure of the category as a whole.
That is, if categories form to maximize the information-rich cluster of
attributes in the environment and, thus, the cue validity or category
resemblance of the attributes of categories, prototypes of categories appear to
form in such a manner as to maximize such dusters and such cue validity still
further within categories.
It is important to note that for natural language categories
both at the superordinate and basic levels the extent to which items have
attributes common to the category was highly negatively correlated with the
extent to which they have attributes belonging to members of contrast
categories. This appears to be part of the structure of real-world categories
It may be that such structure is given by the correlated clusters of attributes
of the real world. Or such structure, may be a result of the human tendency
once a contrast exists to define attributes for contrasting categories so that
the categories will be maximally distinctive. In either case, it is a fact that
both representativeness within a category and distinctiveness from contrast
categories are correlated with prototypicality in real categories. For
artificial categories, either principle alone will produce prototype effects
(Rosch et al., 1976b; Smith & Balzano, personal communication) depending on
the structure of the stimulus set. Thus to perform experiments to try to
distinguish which principle is the one that determines prototype formation and
category processing appears to be an artificial exercise.
Tversky formalizes prototypicality as the member or members
of the category with the highest summed similarity to all members of the
category. This measure, although formally more tractable than that of cue
validity, does not take account, as cue validity does, of an item's
dissimilarity to contrast categories. This issue is discussed further later.
Effects of Prototypicality on
Psychological Dependent Variables
The fact that prototypicality is reliably rated and is
correlated with category structure does not have clear implications for
particular processing models nor for a theory of cognitive representations of
categories (see the introduction to Part III and Chapter 9). What is very clear
from the extant research is that the prototypicality of items within a category
can be shown to affect virtually all of the major dependent variables used as
measures in psychological research.
Speed of Processing: Reaction Time.
The speed with which subjects can judge statements about
category membership is one of the most widely used measures of processing in
semantic memory research within the human information-processing framework.
Subjects typically are required to respond true or false to statements of the
form: X item is a member of Y category, where the dependent variable of
interest is reaction time. In such tasks, for natural language categories,
responses of true are invariably faster for the items that have been rated more
prototypical. Furthermore, Rosch et al. (1976b) had subjects learn artificial
categories where prototypicality was defined structurally for some subjects in
terms of distance of a gestalt configuration from a prototype, for others in
terms of means of attributes, and for still others in terms of family
resemblance between attributes. Factors other than the structure of the
category, such as frequency, were controlled. After learning was completed,
reaction time in a category membership verification task proved to be a
function of structural prototypicality.
Speed of Learning of Artificial Categories (Errors) and
Order of Development in Children.
Rate of learning of new material and the naturally
obtainable measure of learning (combined with maturation) reflected in
developmental order are two of the most pervasive dependent variables in
psychological research. In the artificial categories used by Rosch et al.
(1976b), prototypicality for all three types of stimulus material predicted
speed of teaming of the categories. Developmentally, Anglin (1976) obtained
evidence that young children learn category membership of good examples of
categories before that of poor examples. Using a category-membership
verification technique, Rosch (1973) found that the differences in reaction
time to verify good and poor members were far more extreme for 10-year-old
children than for adults, indicating that the children had learned the category
membership of the prototypical members earlier than that of other members.
Order and Probability of Item Output.
Item output is normally taken to reflect some aspect of
storage, retrieval, or category search. Battig and Montague (1969) provided a
normative study of the probability with which college students listed instances
of superordinate semantic categories. The order is correlated with
prototypicality ratings (Rosch, 1975b). Furthermore, using the artificial
categories in which frequency of experience with all items was controlled,
Rosch et al. (1976b) demonstrated that the most prototypical items were the
first and most frequently produced items when subjects were asked to list the
members of the category.
Effects of Advance Information on Performance: Set,
Priming.
For colors (Rosch, 1975c), for natural superordinate
semantic categories (Rosch, 1975b), and for artificial categories (Rosch et
al., 1976b), it has been shown that degree of prototypicality determines
whether advance information about the category name facilitates or inhibits
responses in a matching task.
The Logic of Natural Language Use of Category Terms: Hedges,
Substitutability into Sentences, Superordination in ASL. Although logic may
treat categories as though membership is all or none, natural languages
themselves possess linguistic mechanisms for coding and coping with gradients
of category membership.
1. Hedges. In English there are qualifying terms such as
"almost" and "virtually," which Lakoff (1972) calls
"hedges." Even those who insist that statements such as "A robin
is a bird" and "A penguin is a bird" are equally true, have to
admit different hedges applicable to statements of category membership. Thus it
is correct to say that a penguin is technically a bird but not that a robin is
technically a bird, because a robin is more than just technically a bird; it is
a real bird, a bird par excellence. Rosch (1975a) showed that when subjects
were given sentence frames such as "X is virtually Y," they reliably
placed the more prototypical member of a pair of items into the referent slot,
a finding which is isomorphic to Tversky's work on asymmetry of similarity
relations (Chapter 4).
2. Substitutability into sentences. The meaning of words is
initimately tied to their use in sentences. Rosch (1977) has shown that
prototypicality ratings for members of superordinate categories predicts the
extent to which the member term is substitutable for the superordinate word in
sentences. Thus, in the sentence "Twenty or so birds often perch on the
telephone wires outside my window and twitter in the morining," the term
"sparrow" may readily be substituted for "bird" but the
result turns ludicrous by substitution of "turkey," an effect which
is not simply a matter of frequency (Rosch, 1975d).
3. Productive superordinates in ASL. Newport and Bellugi
(Chapter 3) demonstrate that when superordinates in ASL are generated by means
of a partial fixed list of category members, those members are the more
prototypical items in the category.
In summary, evidence has been presented that prototypes of
categories are related to the major dependent variables with which
psychological processes are typically measured. What the work summarized does
not tell us, however, is considerably more than it tells us. The pervasiveness
of prototypes in real-world categories and of prototypicality as a variable
indicates that prototypes must have some place in psychological theories of
representation, processing, and learning. However, prototypes themselves do not
constitute any particular model of processes, representations, or learning.
This point is so often misunderstood that it requires discussion:
1. To speak of a prototype at all is simply a convenient
grammatical fiction; what is really referred to are judgments of degree of
prototypicality. Only in some artificial categories is there by definition a
literal single prototype (for example, Posner, Goldsmith, & Welton, 1967;
Reed, 1972; Rosch et al., 1976b). For natural-language categories, to speak of
a single entity that is the prototype is either a gross misunderstanding of the
empirical data or a covert theory of mental representation.
2. Prototypes do not constitute any particular processing
model for categories. For example, in pattern recognition, as Palmer (Chapter
9) points out, a prototype can be described as well by feature lists or
structural descriptions as by templates. And many different types of matching
operations can be conceived for matching to a prototype given any of these
three modes of representation of the prototype. Other cognitive processes
performed on categories such as verifying the membership of an instance in a
category, searching the exemplars of a category for the member with a
particular attribute, or understanding the meaning of a paragraph containing
the category name are not bound to any single process model by the fact that we
may acknowledge prototypes. What the facts about prototypicality do contribute
to processing notions is a constraint - process models should not be
inconsistent with the known facts about prototypes. For example, a model should
not be such as to predict equal verification times for good and bad examples of
categories nor predict completely random search through a category.
3. Prototypes do not constitute a theory of representation
of categories. Although we have suggested elsewhere that it would be reasonable
in light of the basic principles of categorization, if categories were
represented by prototypes that were most representative of the items in the
category and least representative of items outside the category (Rosch &
Mervis, 1975; Rosch, 1977), such a statement remains an unspecified formula until
it is made concrete by inclusion in some specific theory of representation. For
example, different theories of semantic memory can contain the notion of
prototypes in different fashions (Smith, 1978). Prototypes can be represented
either by propositional or image systems (see Chapters 8 and 9). As with
processing models, the facts about prototypes can only constrain, but do not
determine, models of representation. A representation of categories in terms of
conjoined necessary and sufficient attributes alone would probably be incapable
of handling all of the presently known facts, but there are many
representations other than necessary and sufficient attributes that are
possible.
4. Although prototypes must be learned, they do not
constitute any particular theory of category learning. For example, learning of
prototypicality in the types of categories examined in Rosch and Mervis (1975)
could be represented in terms of counting attribute frequency (as in Neuman,
1974), in terms of storage of a set of exemplars to which one later matched the
input (see Chapter 6 and the introduction to Part II), or in terms of explicit
teaching of the prototypes once prototypicality within a category is
established in a culture (e.g., "Now that's a real coat.")
In short, prototypes only constrain but do not specify
representation and process models. In addition, such models further constrain
each other. For example, one could not argue for a frequency count of
attributes in children's learning of prototypes of categories if one had reason
to believe that children's representation of attributes did not allow for
separability and selective attention to each attribute (see Chapter 5 and the
introduction to Part 11).
TWO PROBLEMATICAL ISSUES
The Nature of Perceived Attributes.
The derivations of basic objects and of prototypes from the
basic principles of categorization have depended on the notion of a structure
in the perceived world - bundles of perceived world attributes that formed
natural discontinuities. When the research on basic objects and their
prototypes was initially conceived (Rosch et al., 1976a), I thought of such
attributes as inherent in the real world. Thus, given an organism that had
sensory equipment capable of perceiving attributes such as wings and feathers, it
was a fact in the real world that wings and feathers co-occurred. The state of
knowledge of a person might be ignorant of (or indifferent or inattentive to)
the attributes or might know of the attributes but be ignorant concerning their
correlation. Conversely, a person might know of the attributes and their
correlational structure but exaggerate that structure, turning partial into
complete correlations (as when attributes true only of many members of a
category are thought of as true of all members). However, the environment was
thought to constrain categorizations in that human knowledge could not provide
correlational structure where there was none at all. For purposes of the basic
object experiments, perceived attributes were operationally defined as those
attributes listed by our subjects. Shape was defined as measured by our
computer programs. We thus seemed to have our system grounded comfortably in
the real world.
On contemplation of the nature of many of the attributes
listed by our subjects, however, it appeared that three types of attributes
presented a problem for such a realistic view: (1) some attributes, such as
"seat" for the object "chair," appeared to have names that
showed them not to be meaningful prior to knowledge of the object as chair; (2)
some attributes such as "large" for the object "piano"
seemed to have meaning only in relation to categorization of the object in
terms of a superordinate category - piano is large for furniture but small for
other kinds of objects such as buildings; (3) some attributes such as "you
eat on it" for the object "table" were functional attributes
that seemed to require knowledge about humans, their activities, and the real
world in order to be understood (see Chapter 10). That is, it appeared that the
analysis of objects into attributes was a rather sophisticated activity that
our subjects (and indeed a system of cultural knowledge) might well be
considered to be able to impose only after the development of the category
system.
In fact, the same laws of cognitive economy leading to the
push toward basic level categories and prototypes might also lead to the
definition of attributes of categories such that the categories once given
would appear maximally distinctive from one another and such that the more
prototypical items would appear even more representative of their own and less
representative of contrastive categories. Actually, in the evolution of the
meaning of terms in languages, probably both the constraint of real-world
factors and the construction and reconstruction of attributes are continually
present. Thus, given a particular category system, attributes are defined such
as to make the system appear as logical and economical as possible. However, if
such a system becomes markedly out of phase with real-world constraints, it
will probably tend to evolve to be more in line with those constraints - with
redefinition of attributes ensuing if necessary. Unfortunately, to state the
matter in such a way is to provide no clear place at which we can enter the
system as analytical scientists. What is the unit with which to start our
analysis? Partly in order to find a more basic real-world unit for analysis
than attributes, we have turned our attention to the contexts in which objects
occur - that is, to the culturally defined events in which objects serve as
props.
The Role of Context in Basic-Level Objects and
Prototypes.
It is obvious, even in the absence of controlled
experimentation, that a man about to buy a chair who is standing in a furniture
store surrounded by different chairs among which he must choose will think and
speak about chairs at other than the basic level of "chair."
Similarly, in regard to prototypes, it is obvious that when asked for the most
typical African animal, people of any age will not name the same animal as when
asked for the most typical American pet animal. Because interest in context is
only beginning, it is not yet clear just what experimentally defined contexts
will affect what dependent variables for what categories. But it is
predetermined that there will be context effects for both the level of
abstraction at which an object is considered and for which items are named,
learned, listed, or expected in a category. Does this mean that our findings in
regard to basic levels and prototypes are relevant only to the artificial
situation of the laboratory is which a context is not specified?
Actually, both basic levels and prototypes are, in a sense,
theories about context itself. The basic level of abstraction is that level of
abstraction that is appropriate for using, thinking about, or naming an object
in most situations in which the object occurs (Rosch et al., 1976a). And when a
context is not specified in an experiment, people must contribute their own
context. Presumably, they do not do so randomly. Indeed, it seems likely that,
in the absence of a specified context, subjects assume what they consider the
normal context or situation for occurrence of that object. To make such claims
about categories appears to demand an analysis of the actual events in daily
life in which objects occur.
THE ROLE OF OBJECTS IN EVENTS
The attempt we have made to answer the issues of the origin
of attributes and the role of context has been in terms of the use of objects
in the events of daily human life. The study of events grew out of a interest
in categorizations of the flow of experience. That is, our initial interest was
in the question of whether any of the principles of categorization we had found
useful for understanding concrete objects appeared to apply to the cutting up
of the continuity of experience into the discrete bounded temporal units that
we call events.
Previously, events have been studied primarily from two
perspectives in psychology. Within ecological and social psychology, an
observer records and attempts to segment the stream of another person's
behavior into event sequences (for example, Barker & Wright, 1955; Newtson,
1976). And within the artificial intelligence tradition, Story Understanders
are being constructed that can "comprehend," by means of event
scripts, statements about simple, culturally predictable sequences such as
going to a restaurant (Shank, 1975).
The unit of the event would appear to be a particularly
important unit for analysis. Events stand at the interface between an analysis
of social structure and culture and an analysis of individual psychology. It
may be useful to think of scripts for events as the level of theory at which we
can specify how culture and social structure enter the individual mind. Could
we use events as the basic unit from which to derive an understanding of
objects? Could we view objects as props for the carrying out of events and have
the functions, perceptual attributes, and levels of abstraction of objects fall
out of their role in such events?
Our research to date has been a study rather than an
experiment and more like a pilot study at that. Events were defined neither by
observation of others nor by a priori units for scripts but introspectively in
the following fashion. Students in a seminar on events were asked to choose a
particular evening on which to list the events that they remembered of that
day- e.g., to answer the question what did I do? (or what happened to me?) that
day by means of a list of the names of the events. They were to begin in the
morning. The students were aware of the nature of the inquiry and that the
focus of interest was on the units that they would perceive as the appropriate
units into which to chunk the days' happenings. After completing the list for
that day, they were to do the same sort of lists for events remembered from the
previous day, and thus to continue backwards to preceding days until they could
remember no more day's events. They also listed events for units smaller and larger
than a day: for example, the hour immediately preceding writing and the
previous school quarter.
The results were somewhat encouraging concerning the
tractability of such a means of study. There was considerable agreement on the
kinds of units into which a day should be broken - units such as making coffee,
taking a shower, and going to statistics class. No one used much smaller units:
That is, units such as picking up the toothpaste tube, squeezing toothpaste
onto the brush, etc., never occurred. Nor did people use larger units such as
"got myself out of the house in the morning" or "went to all my
afternoon classes." Furthermore, the units that were listed did not change
in size or type with their recency or remoteness in time to the writing. Thus,
for the time unit of the hour preceding writing, components of events were not
listed. Nor were larger units of time given for a day a week past than for the
day on which the list was composed. Indeed, it was dramatic how, as days
further and further in the past appeared, fewer and fewer events were
remembered although the type of unit for those that were remembered remained
the same. That is, for a day a week past, a student would not say that he now
only remembered getting himself out of the house in the morning (though such
"summarizing" events could be inferred); rather he either did or did
not remember feeding the cat that day (an occurrence that could also be
inferred but for which inference and memory were introspectively clearly
distinguishable). Indeed, it appeared that events such as "all the morning
chores" as a whole do not have a memory representation separate from
memory of doing the individual chores - perhaps in the way that superordinate
categories, such as furniture, do not appear to be imageable per se apart from
imaging individual items in the category. It should be noted that event
boundaries appeared to be marked in a reasonable way by factors such as changes
of the actors participating with ego, changes in the objects ego interacts with,
changes in place, and changes in the type or rate of activity with an object,
and by notable gaps in time between two reported events.
A good candidate for the basic level of abstraction for
events is the type of unit into which the students broke their days. The events
they listed were just those kinds of events for which Shank (1975) has provided
scripts. Scripts of events analyze the event into individual units of action;
these typically occur in a predictable order. For example, the script for going
to a restaurant contains script elements such as entering, going to a table,
ordering, eating, and paying. Some recent research has provided evidence for
the psychological reality of scripts and their elements (Bower, 1976).
Our present concern is with the role of concrete objects in
events. What categories of objects are required to serve as props for events at
the level of abstraction of those listed by the students? In general, we found
that the event name itself combined most readily with superordinate noun
categories; thus, one gets dressed with clothes and needs various kitchen
utensils to make breakfast. When such activities were analyzed into their
script elements, the basic level appeared as the level of abstraction of
objects necessary to script the events; e.g., in getting dressed, one puts on
pants, sweater, and shoes, and in making breakfast, one cooks eggs in a frying
pan.
With respect to prototypes, it appears to be those category
members judged the more prototypical that have attributes that enable them to
fit into the typical and agreed upon script elements. We are presently
collecting normative data on the intersection of common events, the objects
associated with those events and the other sets of events associated with those
objects.2 In addition, object names for eliciting events are varied in level of
abstraction and in known prototypicality in given categories. Initial results
show a similar pattern to that obtained in the earlier research in which it was
found that the more typical members of superordinate categories could replace
the superordinate in sentence frames generated by subjects told to "make
up a sentence" that used the superordinate (Rosch, 1977). That is, the
task of using a given concrete noun in a sentence appears to be an indirect
method of eliciting a statement about the events in which objects play a part;
that indirect method showed clearly that prototypical category members are
those that can play the role in events expected of members of that category.
The use of deviant forms of object names in narratives
accounts for several recently explored effects in the psychological literature.
Substituting object names at other than the basic level within scripts results
in obviously deviant descriptions. Substitution of superordinates produces just
those types of narrative that Bransford and Johnson (1973) have claimed are not
comprehended; for example, "The procedure is actually quite simple. First
you arrange things into different groups. Of course, one pile may be sufficient
[p. 400]." It should be noted in the present context that what Bransford
and Johnson call context cues are actually names of basic-level events (e.g.,
washing clothes) and that one function of hearing the event name is to enable
the reader to translate the superordinate terms into basic-level objects and
actions. Such a translation appears to be a necessary aspect of our ability to
match linguistic descriptions to world knowledge in a way that produces the
"click of comprehension."
On the other hand, substitution of subordinate terms for
basic-level object names in scripts gives the effect of satire or snobbery. For
example, a review (Garis, 1975) of a pretentious novel accused of actually
being about nothing more than brand-name snobbery concludes, "And so,
after putting away my 10 year-old Royal 470 manual and lining up my Mongol
number 3 pencils on my Goldsmith Brothers Formica imitation-wood desk, I slide
into my oversize squirrel-skin L. L. Bean slippers and shuffle off to the
kitchen. There, holding Decades in my trembling right hand, I drop it, plunk,
into my new Sears 20-gallon, celadon-green Permanex trash can [p. 48]."
Analysis of events is still in its initial stages. It is
hoped that further understanding of the functions and attributes of objects
can be derived from such an analysis.
SUMMARY
The first part of this chapter showed how the same
principles of categorization could account for the taxonomic structure of a
category system organized around a basic level and also for the formation of
the categories that occur within this basic level. Thus the principles
described accounted for both the vertical and horizontal structure of category
systems. Four converging operations were employed to establish the claim that
the basic level provides the cornerstone of a taxonomy. The section on
prototypes distinguished the empirical evidence for prototypes as structural
facts about categories from the possible role of prototypes in cognitive
processing, representation, and learning. Then we considered assumptions about
the nature of the attributes of real-world objects and assumptions about
context - insofar as attributes and contexts underlie the claim that there is
structure in the world. Finally, a highly tentative pilot study of attributes
and functions of objects as props in culturally defined events was presented.
REFERENCES
Anglin, J. Les premiers termes de rifirence de 1'enfant. In
S. Ehrlich & E. Tuhring (Eds.), La memoire simantique. Paris: Bulletin de
Psychologie, 1976.
Barker, R., & Wright, H. Midwest and its children.
Evanston, 111.: Row-Peterson, 1955. Battig, W. F., & Montague, W. E.
Category norms for verbal items in 56 categories: A replication and extension
of the Connecticut category norms. Journal of Experimental Psychology Monograph,
1969, 80(3, Pt. 2).
Beach, L. R. Cue probabilism and inference behavior.
Psychological Monographs. 1964, 78, (Whole No. 582). (a)
Beach, L. R. Recognition, assimilation, and identification
of objects. Psychological Monographs, 1964, 78(Whole No. 583). (b)
Berlin, B. Speculations on the growth of ethnobotanical
nomenclature. Language in Society, 1972, 1, 51-86.
Borges, J. L. Other inquisitions 1937-1952. New York:
Washington Square Press, 1%6. Bower, G. Comprehending and recalling stories.
Paper presented as Division 3 presidential address to the American
Psychological Association, Washington, D.C., September 1976.
This work is being done by Elizabeth Kreusi.
Bransford, 1. D., & Johnson, M. K. Considerations of
some problems of comprehension. In W. Chase (Ed.), Visual information
processing. New York: Academic Press, 1973. Bruner, J. S., Olver, R. R., &
Greenfield, P. M. Studies in cognitive growth. New York: Wiley, 1966.
Gar* L. The Margaret Mead of Madison Avenue. Ms., March
1975, pp. 47-48.
Garner, W. R. The processing of information and structure.
New York: Wiley, 1974. Kay, P. Taxonomy and semantic contrast. Language, 1971,
47, 866-887.
Lakoff, G. Hedges: A study in meaning criteria and the logic
of fuzzy concepts. Papers from the eighth regional meeting, Chicago Linguistics
Society. Chicago: University of Chicago Linguistics Department, 1972.
Nelson, K. Concept, word and sentence: Interrelations in
acquisition and development. Psychological Review, 1974, 81, 267-285.
Neuman, P. G. An attribute frequency model for the
abstraction of prototypes. Memory and Cognition, 1974, 2, 241-248.
Newtson, D. Foundations of attribution: The perception of
ongoing behavior. In J. Harvey, W. Ickes, & R. Kidd (Eds.), New directions
in attribution research. Hillsdale, N.J.: Lawrence Erlbaum Associates, 1976.
Pillmer. S. Hierarchical structure in perceptual
representation. Cognitive Psychology, in press.
Peterson, M. J., & Graham, S. E. Visual detection and
visual imagery. Journal of Experimental Psychology, 1974, 103, 509-514.
Piaget, J. The origins of intelligence in children. New
York: International Universities Press, 1952.
Posner, M. 1., Goldsmith, R., & Welton, K. E. Perceived
distance and the classification of distorted patterns. Journal of Experimental
Psychology, 1967, 73, 28-38.
Reed, S. K. Pattern recognition and categorization.
Cognitive Psychology, 1972, 3, 382407.
Rips, L. J., Shoben, E. J., & Smith, E. E. Semantic
distance and the verification of semantic relations. Journal of Verbal Learning
and Verbal Behavior, 1973, 12, 1-20.
Rosch, E. On the internal structure of perceptual and
semantic categories. In T. E. Moore (Ed.), Cognitive development and the
acquisition of language. New York: Academic Press, 1973.
Rosch, E. Linguistic relativity. In A. Silverstein (Ed.),
Human communication: Theoretical perspectives. New York: Halsted Press, 1974.
Rosch, E. Cognitive reference points. Cognitive Psychology,
1975, 7, 532-547. (a)
Rosch, E. Cognitive representations of semantic categories.
Journal of Experimental Psychology: General., 1975, 104, 192-233. (b)
Rosch, E. The nature of mental codes for _color categories
es. Journal of Experimental Psychology. Human Perception and Performance,
1975, 1, 303-322. (c)
Rosch, E. Universals and cultural specifics inhuman categorization.
In R. Brislin, S. Bochner, & W. Lonner (Eds.), Cross-cultural perspectives
on learning. New York: Halsted Press, 1975. (d)
Rosch, E. Human categorization. In N. Warren (Ed.), Advances
in cross-cultural psychology (Vol. 1). London: Academic Press, 1977.
Rosch, E., & Mervis, C. B. Family resemblances: Studies
in the internal structure of categories. Cognitive Psychology, 1975, 7,
573-605.
Rosch, E., Mervis, C. B., Gray, W. D., Johnson, D. M., &
Boyes-Braem, P. Basic objects in natural categories. Cognitive Psychology,
1976, 8, 382-439. (a)
Rosch, E., Simpson, C., & Miller, R. S. Structural bases
of typicality effects. Journal of Experimental Psychology: Human Perception and
Performance. 1976, 2. 491-502. (b)
Shank, R. C. The structure of episodes in memory. In D. G.
Bobrow & A. Collins (Eds.), Representation and understanding: Studies in
cognitive science. New York: Academic Press, 1975.
Smith, E. E. Theories of semantic memory. In W. K. Estes
(Ed.), Handbook of learning and cognitive processes (Vol. 5). Hillsdale, N.J.:
Lawrence Erlbaum Associates, 1978. Smith, E. E., & Balzano, G. J. Personal
communication, April 1977.
Smith, E. E., Balzano, G. J., & Walker, J. H. Nominal,
perceptual, and semantic codes is picture categorization. In J. Cotton & R.
Klatzky (Eds.), Semantic factors in cognition. Hillsdale, N.J.: Lawrence
Erlbaum Associates, 1978.
Tversky, S. Features of similarity. Psychological Review,
1977, 84, 327-352.
Wittgenstein, L. Philosophical investigations. New York:
Macmillan, 1953.
