Barsalou 2019 Ad Hoc Categ

Download as pdf or txt
Download as pdf or txt
You are on page 1of 17

Memory & Cognition

1983, 11 (3), 211-227

Ad hoc categories
LAWRENCE W. BARSALOU
Emory University, Atlanta, Georgia 30322

People construct ad hoc categories to achieve goals. For example, constructing the cate-
gory of "things to sell at a garage sale" can be instrumental to achieving the goal of selling
unwanted possessions. These categories differ from common categories (e.g., "fruit," "fur-
niture") in that ad hoc categories violate the correlational structure of the environment and
are not well established in memory. Regarding the latter property, the category concepts,
concept-to-instance associations, and instance-to-concept associations structuring ad hoc cate-
gories are shown to be much less established in memory than those of common categories.
Regardless of these differences, however, ad hoc categories possess graded structures [i.e.,
typicality gradients) as salient as those structuring common categories. This appears to be
the result of a similarity comparison process that imposes graded structure on any category
regardless of type.

The study of natural categories has been limited (1981) and Smith and Medin (1981), the discovery of
mostly to common categories such as "birds," "furni- graded structure has had a major impact on theories
ture," and "fruit." However, the use of highly specialized of categorization. Graded structure has three aspects.
and unusual sets of items pervades everyday living. Some First, some instances are better examples of a category
examples are "things to take on a camping trip," "pos- than are others; "chair" is a more typical example of
sible costumes to wear to a Halloween party:' and "furniture" than is "bookcase." This aspect of graded
"places to look for antique desks." Since categories like structure has been found in all common categories
these often appear to be created spontaneously for use investigated so far. Rosch (1973, 1975b) found typi-
in specialized contexts, I refer to them as ad hoc cate- cality in color categories (e.g., red, green). Rips, Shoben,
gories. Theories of natural categories primarily reflect and Smith (1973) and Rosch (1973, 1975a) found
what we have learned from common categories. By typicality in common semantic categories (e.g., fruit,
further considering ad hoc categories, we may discover clothing). Rosch and Mervis (1975), Rosch, Simpson,
a more general theory of categorization for which and Miller (1976), Smith, Shoben, and Rips (1974),
common and ad hoc categories are special cases. and Tversky (1977) have since argued that the typicality
This introduction first addresses two central proper- of a category member increases as it becomes more
ties of common categories: graded structure and well similar to other category members. The second aspect
established category representation in memory. A of graded structure is the presence of unclear cases,
comparison-network model is then proposed that items whose category membership is uncertain
accounts for these properties in common categories. (McCloskey & Glucksberg, 1978); people are not sure
The following section shows how this general model can whether "radio" belongs to the category of "furniture."
also account for related predictions in ad hoc categories. The third aspect of graded structure is that the non-
Before going on to four experiments that address these members of a category (i.e., its complement) vary in
predictions, an additional theoretical difference between how similar they are to the concept of the category;
ad hoc and common categories, the degree to which "typewriter" is more similar to the concept of "stereo
they reflect correlational structure, receives brief dis- equipment" than is "dog." This aspect has commonly
cussion. been cited as the reason some false items take longer to
reject in the category verification task than do others
Central Properties of Common Categories (McCloskey & Glucksberg, 1979; Smith et al., 1974);
Graded structure. As noted by Mervis and Rosch "bat" takes longer to reject as a member of "birds"
than does "chair." No evidence bears on whether such
I am particularly grateful to Gordon Bower for supporting
this work and to Herbert Clark for assistance in writing this
similarity gradients cause subjects to reliably rate non-
paper. I am also grateful to colleagues at Stanford and Emory members of a category for typicality within the comple-
universities and to reviewers for helpful discussion and com- ment of the category. But, given the strong relation
ments. The research was supported by Grant MH 13950 from between judged typicality and category verification for
the National Institutes of Mental Health to Gordon Bower and category members, reliable judgments of typicality for
by a National Science Foundation graduate fellowship to the
author. A brief summary of the work supported in this paper nonmembers as well would not be surprising. In sum-
was presented at the American Psychological Association Con- mary, graded structure is a continuum of category
vention in Toronto, 1980. membership, ranging from prototypical members through

211 Copyright 1983 Psychonomic Society, Inc.


212 BARSALOU

unclear cases to prototypical nonmembers (cf. Zadeh, importance of properties and the importance of com-
1965). mon vs. distinctive property sets (cf. Ortony, 1979).
Well established category representation. The second Graded structure in a category results from com-
property of common categories this paper focuses on is puting how similar the concepts for instances, unclear
their possession of well established category represen- cases, and noninstances are to the concept for the
tations. Organization of information to be remembered category. The properties in a category concept are those
is clearly central to recall performance (Bousfield, occurring most often for category instances and least
1953; Mandler, 1967; Tulving, 1962) and is also impor- often for noninstances (see probabilistic concepts in
tant to recognition performance (Mandler, Pearlstone, & Smith & Medin, 1981). The central assumptions are:
Kooprnans, 1969). Taxonomic organization, almost (1) as instances become more similar to a category
exclusively employing common categories, has been one concept, they become more typical of the category;
of the primary types of organization to receive atten- (2) as noninstances become less similar to a category
tion. That taxonomically organized lists are better concept, they become more typical of the category's
recalled than lists of unrelated words has been demon- complement; and (3) the similarity of unclear cases to
strated on numerous occasions (e.g., Bower, Clark, a category concept is close to the minimum amount
Lesgold, & Winzenz, 1969; Cofer, 1967; Puff, 1970). necessary for category membership.
Such organization also results in clustering during recall Rosch and Mervis (1975) report data for common
(e.g., Bousfield, 1953; Bousfield & Cohen, 1953). Most categories consistent with this model. They found that
investigators have viewed the effects of taxonomic the more similar an instance is to all other category
organization as reflecting the existence of preexperi- members (i.e., its family resemblance), the more typical
mental structure in memory. Words from common it is of the category. Assuming that a category concept
categories are easily organized during encoding because is the average (in some sense) of all category members,
memory structures for these categories assimilate pre- how similar an instance is to the category concept
sented information. Retrieval is facilitated during testing
should be at least highly correlated with (if not the same
because these structures provide a network for locating as) how similar the instance is to all other instances.
presented information. When such organization does So, the fmding that typicality correlates highly with
not exist (e.g., for a list of unrelated words), subjects family resemblance is consistent with typicality depend-
have more difficulty organizing and retrieving a list ing on how similar category instances are to their cate-
because (1) subjects have less relevant structure to begingory concept.
with, and (2) the structures used may be created during The spreading activation network. This network
learning and therefore not be well established. Further represents concepts and properties as nodes and repre-
evidence of well established category representations forsents associations between concepts and properties as
common categories comes from the free association pathways that carry spreading activation. A more com-
literature. The highest associates of many words are plex but very similar network model can be found in
often the names of common superordinate categories Collins and Loftus (1975). Loftus (1975) and Rosch
and the names of common contrast categories (e.g., (1975c) provide additional comments of interest to this
"chair" as a cue often produces "furniture" and "table," discussion. In the model I am proposing, each concept
respectively). Superordinates and contrast categories are is associated to properties characteristic of the concept's
high associates presumably because well established referents in the environment. In addition, properties can
memory structure interrelates these categories. be associated to each other and concepts can be associ-
ated to each other. Associations have labels such as
A Comparison-Network Model "has" (e.g., a robin has wings), "cooccurs" (e.g., "gills"
This model, which accounts for graded structures cooccurs with "swims"), and "is an instance of' (e.g.,
and well established category representations in common "robin" is an instance of "bird"). Strength of association
categories, contains two interrelated components: a is free to vary continuously and increases as a function
similarity comparison process and a spreading activation of how frequently and recently an association has been
network. active in working memory. Associations can also be
The similarity comparison process. This process asymmetrical; that is, the strength of association from
computes the similarity of two concepts in working Node X to Node Y may not be the same as the strength
memory. Generally, I will assume that similarity is some from Node Y to Node X. Although activation arrives at
function of the concepts' properties, and specifically, the terminal nodes of both weak and strong associations
I will assume that this function is along the lines of equally fast (Lorch, 1982; Ratcliff & McKoon, 1981),
Tversky's (1977) contrast model. Tversky's account greater activation accumulates at the terminal nodes of
states that two concepts become more similar as the strong associations than of weak ones in a fixed time
number of properties shared by them increases and period. A node is active in working memory when the
the number of distinctive (i.e., nonshared) properties total amount of activation arriving at it is above some
decreases. His model also allows for weighting the threshold value. When a node becomes active, there is
AD HOC CATEGORIES 213

competition for activation among associations leaving Ad Hoc Categories and the
it (i.e., the fan effect; Anderson & Bower, 1973) only Comparison-Network Model
for weak associations. As found by Hayes-Roth (1977), The comparison-network model is sufficiently general
well learned associations do not compete with each to make predictions for ad hoc categories as weIl as to
other for activation. explain the previous findings for common categories.
This network model accounts for the well established Predictions regarding graded structure and category
category representations of common categories in three representation in ad hoc categories are addressed in
ways: Common categories have well established concept- turn.
to-instance associations, well established instance-to- Graded structure. The comparison-network model
concept associations, and well established category predicts that ad hoc categories should exhibit graded
concepts. AIl are well established because of frequent structure. If graded structure results from concepts
and recent processing in working memory. The perfor- bearing different amounts of similarity to a category
mance implications of each type of structure are dis- concept and if concepts vary in similarity from the
cussed in turn. category concepts of ad hoc categories (a safe assump-
(1) Strong concept-to-instance associations in com- tion), then ad hoc categories should exhibit graded
mon categories enable category concepts to easily acti- structure. In contrast, it is possible that ad hoc cate-
vate category instances. Such top-down associations are gories may not be processed in the same way as common
useful when one is trying to generate category members categories. Instead, they may be processed as true
during category production tasks or trying to recall equivalence classes upon which the similarity compari-
information from a categorized list. For example, son process does not operate. Ad hoc categories may be
activating "furniture" might activate "chair," "table," represented more as lists without internal structure than
"desk," and so on. In addition, Barsalou (1981) and as categories possessing typicality gradients. Experi-
Mervis, Catlin, and Rosch (1976) report that typical ment I addresses whether ad hoc categories can possess
instances generally have stronger concept-to-instance graded structure.
associations than atypical instances. As proposed in the next section, an important differ-
(2) Strong instance-to-concept associations in com- ence between ad hoc and common categories is that
mon categories enable instances to activate their cate- ad hoc categories do not have well established category
gory concepts. Such bottom-up associations are useful representations in memory. However, it is hard to see
for categorizing single instances and for organizing how this difference would affect the similarity compari-
multiple instances of the same category. For example, son process, assuming this process is found to generate
perceiving the words "oak," "maple," and "pine" all graded structure in ad hoc categories. Experiments 2a
activate "tree," which can be used for the purposes of and 2b address whether the graded structures of poorly
categorization and organization. established categories (e.g., ad hoc categories) differ
(3) The category concepts for common categories from those of well established categories (e.g., common
are well established in memory because the associations categories ).
between a concept and its properties and between the Category representation. A central difference be-
properties themselves are well established. For example, tween COmmon and ad hoc categories appears to be
"bird" is highly associated to "wings," "flies," "feathers." that common categories have well established category
and so on, which are highly associated among them- representations in memory, whereas ad hoc categories
selves. To the extent a category concept is weIl estab- do not. Ad hoc categories are not wen established simply
lished, it should be easier to locate in memory. This because people rarely, if ever, think of them. This pre-
follows from the assumption that weIl established con- cludes the development and strengthening of associa-
cepts are more "visible" to a memory scanning mech- tions between the nodes representing them. Besides
anism or from the assumption that weIl established observing whether lack of established category repre-
concepts have more pathways into them from other sentation eliminates graded structure in ad hoc cate-
information in memory. gories (as just discussed), the experiments to follow
The similarity comparison process interfaces with explore three ways that this lack of structure may cause
the network in that concepts entering the comparison the processing of ad hoc categories to differ from the
process are concept-property node sets activated above processing of common categories.
threshold. FoIlowing Barsalou (1982), only a subset of The first way lack of category representation may
a concept's properties is usually active. This active subset affect the processing of ad hoc categories centers on
may contain (1) context-independent properties that concept-to-instance associations. These associations
are active on all occasions the concept is processed, and enable a category concept to act as a cue to activate
(2) context-dependent properties that are activated only category instances during category production and
by relevant contexts. For example, "basketball" may recall. Unlike common categories, ad hoc categories may
activate "round" on all occasions, but it may activate not have direct associations from their category con-
"floats" only in contexts involving bodies of water. cepts to their instances. If so, retrieving instances from
214 BARSALOU

ad hoc categories should be more difficult than retriev- "dog," "stereo," and "blanket." This category's instances
ing instances from common categories. Experiments 2a do not appear to share correlated properties. More-
and 3 explore this prediction. over, its instances share many correlated properties with
The second way lack of category representation may entities in the complement of the category. For exam-
affect the processing of ad hoc categories centers on ple, "dog" also belongs to the common category of
instance-to-concept associations. These associations "mammals," many of which belong to the complement
enable people to activate categories to which an instance of "things to take from one's home during a fire."
belongs. Unlike common categories, ad hoc categories If ad hoc categories cut across the correlational
may not have direct associations from their instances to structure of the environment, then why do people per-
their category concepts. If so, categorizing instances into ceive them as categories? Most likely, this is because
ad hoc categories should be more difficult than cate- ad hoc categories are instrumental to achieving goals. For
gorizing them into common categories. Experiment 4 example, someone trying to escape a burning home and
explores this prediction. minimize loss might try to construct the category of
The third way lack of category representation may "things to take from one's home during a fire" before
affect the processing of ad hoc categories centers on how heading for safety. Similarly, someone interested in tak-
well established their category concepts are. Unlike ing a trip would need to consider "things to pack in a suit-
common categories, the properties composing ad hoc case." To the extent that someone is achieving either of
category concepts may not be well associated, since these goals for the first time, the corresponding cate-
these properties have rarely, if ever, been processed gories should not have well established category represen-
simultaneously. If so, the category concepts of ad hoc tations in memory. As noted in the general discussion,
categories should not be as accessible as those of com- however, frequently used ad hoc categories may develop
mon categories. Experiment 3 explores this prediction. well established category representations much like
those of common categories.
The Relation of Ad Hoc and Common Before proceeding to experiments that address graded
Categories to Correlational Structure structure and category representation in ad hoc cate-
An important difference between ad hoc and com- gories, it is necessary to comment on the sampling of
mon categories not addressed empirically in this paper, ad hoc categories. Ad hoc categories will be defined for
but certainly worthy of future attention, centers on the the purpose of this paper as sets that (1) violate corre-
correlational structure of the environment. As Rosch, lational structure and (2) are usually not thought of by
Mervis, Gray, Johnson, and Boyes-Braem (1976) have most people. Clearly, there are an indefinitely large
noted, properties of entities in the environment are not number of such categories, and it would probably be
independent but, instead, form clusters of correlated impossible to enumerate them all. At this point, it is
properties. For example, if an entity has feathers, there even difficult to imagine what kinds of ad hoc categories
is a much higher probability that it flies and builds exist. For this reason, these experiments do not attempt
nests than that it swims and has gills. Entities instantiat- to draw conclusions about all ad hoc categories but
ing a set of correlated properties are very similar to each concentrate instead on the nature of a few of them.
other and are very different from entities instantiating Although these experiments can primarily be interpreted
other sets of correlated properties. For example, differ- with respect to existential as opposed to universal
ent kinds of birds are very similar to each other and are claims, they provide constraints on a general theory of
very different from members of fish and vehicles. categorization. Finding that some uncommon cate-
Rosch, Mervis, Gray, Johnson, and Boyes-Braem (1976) gories exhibit graded structure forces a general theory
show that people are sensitive to the correlational to include a mechanism that can generate such structure
structure of the environment and that they prefer to use in a wider range of categories. Finding that categories
categories that take maximal advantage of it. vary in how well established they are forces a general
What I have been referring to as common categories theory to include mechanisms that account for this
appear to reflect correlatonal structure. They circum- difference. And finding that graded structures occur in
scribe sets of entities that share many correlated prop- poorly established categories further constrains the
erties and that do not share many properties with mem- mechanism responsible for generating graded structure.
bers of other categories (Rosch & Mervis, 1975). We
readily perceive these sets as categories because they EXPERIMENT 1
have so much in common (although, as Rosch and
Mervis show, no property or correlation of properties Rather than possessing graded structure, ad hoc
need be true of all a category's members). categories may be represented as unordered lists in
In contrast, ad hoc categories appear to violate the which all instances are equally good members. Alterna-
correlational structure of the environment. Consider the tively, the same similarity comparison process that
category of "things to take from one's home during a generates graded structure for common categories may
fire," which contains members as diverse as "children," also operate during the processing of ad hoc categories.
AD HOC CATEGORIES 215

This experiment and the next investigate whether ad hoc Table I


categories possess graded structure. Do subjects reliably Examples of Materials Used in Experiments I and 4
perceive typicality gradients in these categories? Are AD HOC CATEGORIES
subjects uncertain regarding the membership of certain
Ways to Make Friends
instances, namely, unclear cases? And do subjects Martin had moved from the midwest to the west coast over a
reliably perceive some nonmembers to be more typical year ago. He had encountered much trouble making friends
nonmembers than others? since he had arrived in California and could not think of anyone
he presently considered a good friend. He decided it was time to
do something about it.
Method
Design. Subjects performed two categorizing operations on Experiment I Item Set
eight sets of six items. For each set, they were asked first to join a card playing club
separate out those items that belonged to some category and, get convicted for murder
second, to rank order all six items for how good a member each don't take a bath more often than once a month
was of that category. go back to school
Materials. Each subject received a booklet containing instruc- have a garage sale
tions followed by eight pages of category materials, one for each get convicted for burglary
category. Each page contained a context vignette, the name of Experiment 4 Item Set
an ad hoc category, and six randomly ordered labels in a column,
some of these denoting instances of the category. The categories get involved in local politics
were "things to inventory at a department store," "ways to get rich
make friends," "things that conquerors take as plunder," have a garage sale
"nouns," "ways to escape being killed by the Mafia," "things go to school
that babies do," "times to write a term paper," and "things Ways to Escape Being Killed by the Mafia
that could fall on your head." Roy was in big trouble. The Mafia had a contract out on
Each context vignette described a person engaged in an him for double-crossing them. He knew he couldn't continue
activity. The category label that followed denoted a category living in Las Vegas or he'd be dead in a week. So he started
relevant to the person's goals. The vignettes were used to estab- thinking quickly about alternatives.
lish the ad hoc categories in goal contexts. For each of four item
sets, three items were obviously from the a priori category and Experiment I Item Set
the other three were not. For each of the other four sets, two change your identity and move to the
items were obviously from the a priori category, two were not, mountains of South America
and the remaining two were, intuitively, unclear cases. Table I move to the remote reaches of Wyoming*
contains examples of the contexts and item sets. stay where you're presently living in Las Vegas
The instruction sheet directed subjects to read both the move to Reno*
vignette and the category label, to look through the six items move to the mountains of Mexico
that followed, and to circle those belonging to the category; change where you're living in Las Vegas
there was no constraint on the number they could circle. Next, Experiment 4 Item Set
they were to rank all six items from the best example of the
category to the worst, with no ties. move to the remote reaches of Wyoming
Subjects and Procedure. Twelve undergraduates participated sail around the world
for either course credit or pay. Half received one randomized go to Mexico
version of the list, and the other half received a different ran- become a drunk in Detroit
domized version. Subjects worked through the booklets at their
own pace and had as much time as necessary to complete the COMMON CATEGORIES
experiment. Fruit
Dan thoroughly enjoyed food. His favorite time of the year
was summer because of the abundance of fresh food that was
Results available.
Unclear cases. Agreement for category membership
Experiment 4 Item Set
for a given item in an item set was determined as fol-
apple
lows. If a subject had circled an item, it was scored as
orange
+1; if the subject had not, it was scored as -1. The banana
scores for the item were then summed across subjects. If peach
all 12 subjects circled the item, it received a score of
+12; if no subject circled the item, it received a score of RANDOM CATEGORIES
- I 2; if half the subjects circled the item, it received a Horace was designing a computer system that would operate
score of O. The absolute value of the score was divided the traffic signal system in a major urban area. He had to find a
by 12. This measure ranged from 0 to 1, 0 indicating competent group of programmers to help him do the program-
ming for the system.
no agreement (for an unclear case) and I indicating
complete agreement. The absolute value was taken to Experiment 4 Item Set
measure agreement for nonmembers as well as for memo blue
erase
bers of the category. The overall agreement for an item riddle
set was simply the average of the agreement scores for monkey
the six items in that set.
Note-Only ad hoc categories were used in Experiment 1. The
It should be noted that the average number of items c.ategory labels were not presented in Experiment 4.
circled per item set was 3.15, with subjects circling "Unclear case.
216 BARSALOU

3 items 70% of the time and circling 2-4 items 95% of EXPERIMENTS 2A AND 28
the time. It can be easily shown that the overall agree-
ment score should be .04 if subjects were guessing and As shown by Experiment 1, ad hoc categories can
had circled 3.15 items on the average.' possess graded structure. But how do these structures
The mean agreement across item sets for this analysis compare with those found in common categories? To
was .88. The average agreement for item sets not having what extent might well established category represen-
unclear cases was .97, compared to .79 for item sets tations in common categories cause their graded struc-
with unclear cases (Mann-Whitney U =0, p = .014). tures to differ from those of ad hoc categories?
An analysis of the item sets having unclear cases showed As mentioned earlier, salient typicality gradients have
that lower agreement for these sets was entirely the always been found in common categories. In addition,
result of poor agreement for five of the eight a priori common categories exhibit another form of graded
unclear cases; the average for these five items was .17, structure: When subjects generate instances to a category
compared to .96 for all other items from these four sets. name, instances vary in how frequently they are gene-
This demonstrates the presence of unclear cases in rated; some instances are better examples during genera-
ad hoc categories; namely, subjects can be divided about tion than others. These two forms of graded structure,
whether certain items are members of ad hoc categories.? typicality and production frequency, are well correlated
Graded structure. Agreement for graded structure was for common categories (Barsalou, 1981; Mervis et al.,
determined by computing Kendall's coefficient of con- 1976).
cordance across the subjects by items matrix of rankings Experiments 2a and 2b compare these two forms of
for each item set. A transformation of this statistic graded structure for common and ad hoc categories.
(Guilford & Fruchter, 1973) estimates how much a Experiment 2a compares the distributions of responses
given subject agrees with every other subject on the given to both category types during exemplar genera-
average. It is an estimate of subject agreement and not tion. In particular, are some ad hoc category instances
the stability of an item set's means. If ad hoc categories more dominant than others during generation, as is the
do not possess salient graded structures, then subjects case for common categories? If so, are the most domi-
should show no agreement, and this statistic should nant responses for ad hoc categories as dominant as the
approach O. To the extent there is a salient graded most dominant responses for common categories?
structure perceived by all subjects, this statistic should Experiment 2b compares typicality gradients for the two
approach 1. Across categories, the average agreement category types. In particular, do subjects agree as much
was .87. Subjects' high agreement demonstrates ad hoc in their typicality judgments for ad hoc categories as
categories possess salient graded structure. they do for common categories? Also, do ad hoc cate-
Agreement was also computed for clear category gory instances vary as much in typicality as common
members alone and for clear category nonmembers alone. category instances do? That is, are typicality gradients
Average agreement was .54 for category members and in ad hoc categories as "steep" as those in common
.37 for category nonmembers, revealing both internal categories?
and external graded structure. Agreement for internal The comparison-network model predicts that typi-
structure was greater than for external structure in only cality gradients, as indexed by typicality judgments,
three of the eight item sets. should be very similar for ad hoc and common cate-
The presence of unclear cases did not affect subjects' gories. This follows from the assumptions that (1) the
typicality rankings. The average agreement was .92 for same similarity comparison process constructs graded
item sets having unclear cases and .88 for those without; structure for both category types and (2) how well
this difference was not significant (Mann-Whitney established a category is in memory does not affect this
U = 4, P = .171). So, unclear cases led to less agreement process. In contrast, the model predicts that subjects
for category membership but not for typicality. Subjects should show less agreement when generating category
may agree on the graded structure underlying a category members for ad hoc than for common categories. This
but be uncertain where to draw the category boundary follows from the assumption that common categories
along this continuum. have well established concept-to-instance associations to
their typical instances, whereas ad hoc categories do not.
Discussion Consequently, most subjects should access the typical
Ad hoc categories can possess graded structure. Sub- instances of the common categories and show high
jects showed high agreement for overall graded structure agreement, whereas subjects' search of ad hoc categories
in general and for internal and external graded structure should be more random and, therefore, show less agree-
in particular. Further evidence of graded structure in ment.
ad hoc categories stems from the presence of unclear
Method: Experiment 2a
cases. Excluding unclear cases, however, subjects were in Twenty subjects generated the first four exemplars that came
excellent agreement over which items were and were not to mind for nine common and nine ad hoc categories. Half the
category members." subjects received the 18 categories in one random order, and half
AD HOC CATEGORIES 217

received them in another. Subjects were introductory psy- conclusion that exemplar production is more consistent
chology students participating to receive course credit. The for common than for ad hoc categories. Ad hoc cate-
common categories were 9 of the 10 common categories used by gories, however, clearly show graded structure in the
Rosch (1975a): "birds," "sports," "fruit," "weapons," "vege-
tables," "vehicles," "clothing," "furniture," and "tools." The sense that some instances are more dominant than others.
ad hoc categories were what intuitively appeared to be atypical
and infrequently used categories. Four of these were drawn from Method: Experiment 2b
the previous experiment: "ways to make friends," "ways to For each of the 18 categories in Experiment 2a, the exemplars
escape being killed by the Mafia," "things that can fall on your were rank ordered by production frequency (i.e., the number of
head," and "plunder taken by conquerors." The other five were subjects, from I to 20, generating an item). Six exemplars were
"things that can be walked upon," "things that can float," then selected from each category, one from the highest level of
"things that are poisonous," "things that can attack something," production frequency, one from the lowest level (always I),
and "things that have a smell." and the remaining four from, as much as possible, equally spaced
intervals between the highest and lowest levels. Two versions of
the stimuli were constructed. Each had the 18 categories in a
Results: Experiment 2a
different random order; within each version, the six exemplars
Within the 80 exemplars generated for each category, from the same category were in a different random order. The
the average number of different exemplars per category six exemplars appeared in a column below their category label.
was 20.33 for the common categories and 37.00 for the The instructions defined typicality and directed subjects to
ad hoc categories [t(16) = 6.54, P < .001]. The distribu- judge the typicality of each exemplar with respect to its cate-
gory label. Twelve subjects rated the exemplars' typicality on a
tions for average number of items per category as a scale from 1 to 7, where 1 referred to a most unusual exemplar
function of average number of subjects generating an and 7 to one of the best. Twelve other subjects ranked the
item are shown in Table 2. The functions for both the exemplars for typicality; no ties were allowed. Six subjects in
common and ad hoc categories were generally decreas- each task received each of the two list versions. Subjects were
ing and negatively accelerated. However, there was an 24 introductory psychology students participating to receive
course credit.
increase in frequency for the common categories for
items generated by 16-20 subjects. Also, the function
Results: Experiment 2b
for the ad hoc categories decreased more rapidly than
Ratings. For each of the 18 categories, the intraclass
the function for the common categories. The average
correlation coefficient (Guilford & Fruchter, 1973)
proportion of subjects producing the ith most generated
was computed for the subjects by items matrix of rat-
item per category is shown in Table 3 (for i = 1-5).
ings. This coefficient can be interpreted as measuring
For each of the five levels shown, the proportion is
how much subjects agree on the typicality of a cate-
higher for the common categories, this being significant
gory's exemplars. Specifically, the coefficient estimates
by a sign test (p < .05). These three results-the average
how much a given subject's ratings correlate with
number of different items generated, the characteristics
another's on the average; it does not measure how stable
of the frequency as a function of agreement distribu-
the mean for each item is. Its average value across cate-
tions, and the proportion of subjects generating the five
gories within category types is shown in Table 4. Sub-
most generated items-are all in agreement with the
jects agreed equally well for both category types [t(16}
= .94, P > .20] .
Table 2 The average level of typicality differed marginally for
Effect of Category Type on Production Frequency Distribution: common and ad hoc categories [t(16) = 1.93, .1 0 >
Average Number of Items per Category (Experiment 2a)
p > .05; see Table 4]. The standard deviation of the
Category Type mean typicality ratings for each category's items was
N Common Ad Hoc computed, and there was no difference between cate-
gory types [t(16} =1.34, p > .20; see Table 4]. In fact,
1-5 15.77 34.00
6-10 2.56 2.21
the average standard deviation for the ad hoc categories
11-15 .77 .77 was slightly larger than that for the common categories.
16-20 1.21 .00 This latter finding indicates that the range of typicality
Note-N = number of subjects (of 20) generating an item.
values for items in common and ad hoc categories was
comparable. This also indicates that a difference in range
did not bias the comparison between intraclass correla-
Table 3
tions for the two category types. If the ranges for one
Average Proportion of SUbjects Producing the ith Most
Generated Item per Category (Experiment 2a) category type had been less on the average, this could
have relatively reduced the correlations for that type.
Category Type
ith Most Rankings, For each of the 18 categories, a transfor-
Generated Item Common Ad Hoc mation of Kendall's coefficient of concordance was
1 .84 .56 computed for the subjects by items matrix of rankings
2 .69 .39 (Guilford & Fruchter, 1973). This statistic, like the
3 .46 .31 intrac1ass correlation, can be interpreted as measuring how
4 .36 .26 much subjects agree on the typicality of a category's
5 .27 .20
exemplars. Unlike the intraclass correlation, Kendall's
218 BARSALOU

Table 4 typical members such that these instances become


Summary Data for Experiment 2b particularly well associated to their categories.
Category Type To provide further support for the claim that concept-
to-instance associations are better established in com-
Measure Common Ad Hoc
mon categories than in ad hoc categories, additional
Ratings data were collected. Twelve new subjects generated
Intraclass Correlation .50 .56 exemplars to the nine common categories and seven of
Typicality 5.08 4.68 the ad hoc categories for a fixed time period (two
Within-Category SD 1.44 1.65
ad hoc categories, "ways to make friends" and "ways
Rankings to keep from being killed by the Mafia," could not be
Coefficient of Concordance .57 .54 used, since their exemplars are typically described by
Note-All measures are averaged across categories. more than one word). A tape recorder presented the
category names to the subjects intermixed in one of two
coefficient does not have the problem of possible random orders. For 15 sec following the end of each
differences in range. This is because the standard devia- name, subjects wrote down as many category instances
tion of a subject's rankings within a set is constant. The as they could think of. Subjects generated 5.67 exemplars
average value of this statistic across categories within on the average for the common categories and 4.22
category types is shown in Table 4. Again, subjects exemplars for the ad hoc categories [t(14) = 5.58,
agreed equally well for both category types [t( 16) = .36, p < .001]. There was no overlap between the distribu-
p > .50]. tions of mean exemplars generated per category for the
Intenneasure correlations. The following six correla- two category types. Stronger concept-to-instance associ-
tions across item averages are all significant at the ations for common than for ad hoc categories enabled
a = .001 level. Typicality ratings correlated with typi- faster access to the instances of common categories.
cality rankings .89 for the common categories and .89 There is an alternative explanation for the differences
for the ad hoc categories. The ratings correlated with between ad hoc and common categories in production
production frequency (from Experiment 2a) .80 for the consistency and access time: These ad hoc categories
common categories and .66 for the ad hoc categories; may have contained more exemplars than the common
the difference between these latter two correlations categories. Regarding consistency, there may simply
approached significance (z = 1.55, p= .12). The rankings have been a lower probability of sampling an item from
correlated with production frequency .90 for the com- an ad hoc than from a common category. With respect
mon categories and .65 for the ad hoc categories; the to access time, greater numbers of exemplars may lead
difference between these two correlations was signifi- to more interference and, therefore, slower access.
cant (z = 3.45, p < .001). This difference suggests that Although both familiarity and category size may affect
high- and low-typicality instances of common cate- exemplar production, two of the present results provide
gories may differ more in how strongly they are assoc- evidence for the familiarity explanation. (1) In Table 3,
iated to their category concepts than do high- and low- the number of exemplars generated for common cate-
typicality instances of ad hoc categories. gories by 16-20 subjects was greater than the number
generated by 11-15 subjects. This could not be the result
Discussion of a difference in sampling probabilities, since the sam-
As predicted by the comparison-network model, pling explanation predicts frequency to be a monotonic
typicality gradients derived from typicality ratings and decreasing function of the number of subjects. Rather,
rankings were equally salient for the ad hoc and com- this "bump" appears to be the result of familiarity with
mon categories. This suggests, first, that subjects use the the most prototypical exemplars. Notably, no such
same similarity comparison process to construct graded bump occurs for the ad hoc categories. (2) The correla-
structure for both category types and, second, that this tions between typicality and production frequency were
process is not affected by how well established a cate- higher for common than for ad hoc categories. This
gory is in memory. suggests that the high- and low-typicality instances of
Similar to common categories, ad hoc categories also common categories differ more than those of common
showed graded structure as indexed by production categories in the strength of their concept-to-instance
frequency. Some members of ad hoc categories are more associations. If so, then the high-typicality instances of
dominant during exemplar generation than others. As common categories may well have stronger concept-to-
predicted by the comparison-network model, however, instance associations than the high-typicality instances
there was less consistency in exemplar production for of ad hoc categories. Experiments 3 and 4 provide
ad hoc than for common categories. This suggests that further evidence that common categories are much
the category types differ in the extent to which their better established in memory than are ad hoc categories.
concept-to-instance associations are established in It is of interest to note that robust typicality gradients
memory. People have more experience with common occurred in categories defmed by necessary and suf-
categories and, therefore, establish stronger associations ficient conditions. Many of the ad hoc categories were of
to these exemplars. Also, cultural forces may focus on the form "things that exhibit X," in which X was a
AD HOC CATEGORIES 219

necessary and sufficient condition (e.g., things that have quently, common categories should be accessed more
a smell, things that can float). Thus, unequal category often than ad hoc categories. In addition, if ad hoc
membership exists even in well defined categories. This categories are accessed no more frequently than random
could be the result of either of two factors. First, a categories, this would suggest that the concepts for
necessary and sufficient condition may be possessed by ad hoc categories are indeed not well established in
exemplars to varying degrees, in different manners, or memory.
with different frequencies. "Milk" and "coffee" vary in It should be noted that the exemplars of the common,
degree of "smell"; "perfume" and "skunk" vary in man- ad hoc, and random categories were made equally acces-
ner of "smell"; and "basketball" and "sailboat" vary in sible in this experiment by equating their word fre-
frequency of "floats." Such differences in possession of quency and imageabliity. Any differences observed in
category criteria may be correlated with typicality. category access, therefore, can be attributed to differ-
Seocnd, irrelevant properties associated with a necessary ences in how well established the category concepts are
and sufficient condition may enter into typicality judg- for each category type, but not to differences in the
ments. The criteria for "medical doctor" are well defmed, accessibility of their exemplars.
but someone possessing irrelevant properties of the Exemplar retrieval. Two predictions follow from the
doctor-personality stereotype may be a better example hypothesis that concept-to-instance associations are
of "doctor" than someone who does not. Barsalou better established for common than for ad hoc cate-
(1981) offers direct evidence that these factors deter- gories. First, more correct exemplars should be retrieved
mine typicality in categories defined by necessary and from common than from ad hoc categories for accessed
sufficient conditions. categories. Second, there should be a higher intrusion
rate for common categories. Both predictions follow
EXPERIMENT 3 from the assumption that stronger concept-to-instance
associations result in instances more likely being acti-
The production frequency results of Experiment 2a vated above threshold during retrieval. Increased avail-
suggest that concept-to-instance associations are better ability of instances via these associations may also
established in memory for common than for ad hoc facilitate organization at encoding.
categories. The current experiment tests this hypothesis
more directly and also tests the hypothesis that the Method
category concepts of common categories are better Design. Three groups of subjects each received 48 words
partitioned into 12 clusters of 4 words each. One group received
established than are those of ad hoc categories. One a list of common category exemplars blocked by category, a
group of subjects received words from common cate- second group received a list of ad hoc category exemplars blocked
gories, a second group received words from ad hoc by category, and a third group received a list of clusters contain-
categories, and a third group received clusters of words ing unrelated words. Half the subjects for each category type
that were unrelated. For the common and ad hoc received the word sets in one random order, and half received
them in another. Within each order, the words for the same set
category lists, a category's instances were grouped and appeared in a different random order. Six subjects were nested
preceded by their category label. All subjects performed in each of the category type by order cells of the design.
a free recall of the words, and the common and ad hoc All subjects performed a free recall of the list after it was
subjects performed a subsequent cued recall. presented. Subjects in the common and ad hoc category condi-
tions were then given the category labels as cues for further
Two measures of recall from categorized lists are of recall. (Random category subjects could not perform a cued
interest (see Tulving & Pearlstone, 1966): (1) the num- recall, since their list did not contain category labels.) In the
ber of categories accessed during recall (a category is common and ad hoc category conditions, half the subjects in
accessed if at least one of its exemplars is recalled), each category type by order cell received the cues in one random
and (2) the average number of exemplars retrieved per order, and half received them in another. Before learning the
critical lists, all subjects received and free recalled the same
accessed category. The predictions of the comparison- practice list, which contained 36 unrelated words similar to
network model are as follows. those in the critical lists.
Category access. If common categories have better Materials. The common categories, drawn from Battig and
established category concepts than ad hoc categories, Montague (1969) and Rosch (1975a), were "furniture," "cloth-
ing," "vehicles," "birds," "sports," "fruit," "vegetables,"
then the category concepts for common categories "insects," "trees," "animals," "musical instruments," and
should be accessed more easily during free recall. This is "colors." The ad hoc categories, which seemed to be atypical
in accordance with the well known fact that high- and infrequently used, were "where eating can occur," "can fall
frequency responses are better recalled than are low- on your head," "has a smell," "can be walked upon," "plunder
frequency responses (Hall, 1954). It can be explained taken by conquerors," "can attack something," "can be used for
hitting," "manufactured by humans," "is poisonous," "can be
by assuming that better established concepts are more eaten," "is a liquid," and "can float." All the exemplars were
"visible" to a scanning process or have more pathways common single-word nouns. Exemplars from the common and
into them from other information in memory. A con- ad hoc categories were typical of their respective categories.
sequence of category concepts being better established Those for "vegetables" were "potato," "corn," "celery," and
for common than for ad hoc categories is that it should "spinach"; those for "manufactured by humans" were "tele-
phone," "helicopter," "camera," and "refrigerator"; those for
be easier to retrieve the exemplars of common cate- one of the random categories were "grease," "spider," "admiral,"
gories during recall via their category concepts. Conse- and "copper."
220 BARSALOU

The three lists were equated for Kucera and Francis (967) computed from a one-way ANOVA performed across the
frequency (21.96, 22.52, and 21.19 for the common, ad hoc, free recall data for all three category types. Except
and random categories, respectively; F < 1). To insure that the
stimulus sets were equivalent in imagery, 12 undergraduates
where noted, each ANOVA was performed twice for
rated how easy it was to image the referent of each word in the each dependent measure: once across subject averages
three lists. On a 1-7 scale, with 7 meaning highest imagery, the and once across item (i.e., category) averages. The
average ratings were 5.42, 5.32, 5.44, for the common, ad hoc, results from the subject and item analyses were then
and random categories, respectively (F < 1). combined to perform min F', planned comparisons of
Subjects and Procedure. The subjects, 36 undergraduates
participating to receive course credit, were told they would be interest (Clark, 1973). Analyses of the intrusion data
presented with two lists to learn for free recall. The practice were performed on frequencies; all other analyses were
list was presented by a slide projector, one word at a time at a performed on proportions transformed using arcsins,
2.5-sec rate, followed by a buffer task lasting 3 min. Subjects as suggested by Winer (1971). The subject means from
were then asked to recall, for 5 min, as many words as possible
in any order from the list.
all analyses are shown in Table 5. There was no differ-
Subjects were then informed about the categorized list to ence between the three list conditions on the practice
come. Subjects in the common and ad hoc category conditions list [min F'(2,62) = 1.60, P > .10] .
were told not to recall the category labels, but that these would Overall exemplar recall. Common category subjects
be useful in organizing and learning their list. Subjects in the recalled more exemplars than did ad hoc category sub-
random category condition were told to learn each cluster of
four words as a group. All subjects were told about the format jects during free recall [min F'(l,39) = 16.72, r < .001]
of their list, which was then presented by a slide projector, one and also during cued recall [min F'(1,39) = 13.40,
item at a time at a 2.5-sec rate. For the common and ad hoc P < .00 1]. Cued recall was superior to free recall both
category conditions, a category label was presented, followed by for common category subjects [min F'(1,38) = 14.70,
the individual presentation of each exemplar. After presentation
of the fourth exemplar, the label of the next category was pre-
P < .001] and for ad hoc category subjects [min F'(l ,39)
sented and followed by the individual presentation of its exem- = 18.18, P < .00l). There was no Recall Type by Cate-
plars, this cycle continuing until all categories had been pre- gory Type interaction (min F' < 1). Random category
sented. For the random category list, a string of Xs appeared in subjects recalled fewer exemplars during recall than
place of the category labels. did common category subjects [min F'(1,64) =8.65,
P < .01]; the difference between random category
After working on a buffer task for 3 min, subjects attempted
to recall, for 5 min, as many exemplars as possible in any order
from the second list only; they did not recall the category and ad hoc category subjects was not significant
labels. Subjects in the common and ad hoc category conditions [min F'(1,62) = 1.16, p> .25], although it favored the
were then given the category labels and again asked to recall, ad hoc category subjects.
for 5 min, as many words as possible in any order from the list.
Category access. Common category subjects accessed
more categories than did ad hoc category subjects during
Results free recall [min F'(1,36) = 9.51, p < .01] . There was no
There are four dependent measures of interest. "Over- significant difference for cued recall [min F' (1,38) =
all exemplar recall" refers to the proportion of 48 cate- 2.73, p > .10] , although it was in the same direction as
gory exemplars recalled by a subject. "Category access" that for free recall. No difference was expected, since
refers to the number of categories for which a subject providing cues should have equalized accessibility.
recalled at least one of the four exemplars. "Exemplars More categories were accessed during cued recall than
per category recall" refers to the average number of during free recall both for common category subjects
exemplars per category recalled by a subject (for only [min F'(1,36) = 15.14, P < .001] and ad hoc category
those categories in which one or more exemplars were subjects [min F'(1,35) = 28.29, p < .00l). There was no
recalled). "Intrusions" refers to the total number of Recall Type by Category Type interaction [min F'(l ,33)
intrusions made by a subject. = 1.01, p > .25]. The difference between random cate-
For each measure, a two-way ANOVA was per- gory subjects and common category subjects for free
formed on the free and cued recall data for the common recall approached significance [min F'(1,62) = 3.40,
and ad hoc category subjects. Comparisons involving the .10> p > .05]. Notably, there was no difference
free recall data for the random category subjects were between random category subjects and ad hoc category

Table 5
Effects of Category Type on Free and Cued Recall (Experiment 3)
Category Type
Free Recall Cued Recall
Measure Random Ad Hoc Common Ad Hoc Common
Overall Proportion of Exemplars Recalled .33 .42 .58 .59 .73
Number of Accessed Categories (of 12) 7.33 7.33 9.42 10.75 11.75
Average Number of Exemplars/ Accessed Category (of 4) 2.15 2.72 2.94 2.59 2.96
Total Intrusions 1.08 .92 2.42 1.44 3.50
Note-All measures are the average per subject.
AD HOC CATEGORIES 221

subjects (min F' < I). In fact, the number of categories .04 [t(IO) = .13, p > .50] and -.16 [t(10) = .52,
accessed was identical. p> .50], respectively. Thus, label length does not
Exemplars per category recall. Common category appear to have been a factor. Furthermore, subjects were
subjects retrieved more exemplars per accessed category instructed not to recall the category labels, but only to
than did ad hoc category subjects, both during free use them for organizing the word sets. This alternative
recall [min F'(l,37) = 5.30, P < .05] and during cued explanation has no bearing on exemplar recall, which
recall [min F'(l,42) = 18.64, P < .001]. There was no was computed only for accessed categories.
difference between cued recall and free recall for the Ad hoc categories have some capability as mnemonic
common category subjects (min F' < I) or for the devices. This was evidenced by ad hoc categories show-
ad hoc category subjects [min F'(1,43) = 2.35, p > .10]. ing greater exemplar recall per accessed category than
There was no Recall Type by Category Type interaction random categories. Although ad hoc categories are not
[min F'(1 ,44) = 1.85, p > .10]. Random category sub- well, if at all, represented in memory, they are at least
jects during free recall retrieved fewer exemplars per able to provide organizational schemes for presented
accessed category than did common category subjects information. The concepts for these categories, however,
[min F(I,65) = 7.67, P < .01] and ad hoc category were no more accessible than the representations of
subjects [min F'(1,65) = 4.04, P < .05]. random groups of words. In fact, the average numbers of
Intrusions. Common category subjects produced categories accessed for these two category types were
more intrusions than did ad hoc category subjects, both identical, 7.33 and 7.33.
during free recall [min F'(1,43) =7.63, p<.OI] and
during cued recall [min F'(1,43) = 14.72, P < .001]. EXPERIMENT 4
The difference between cued recall and free recall was
marginally significant for common category subjects The previous two experiments show that category
[min F'(1,43) = 3.98, .10 > P > .05], but not for ad hoc concepts and concept-to-instance associations are better
category subjects (min F' < I). There was no Recall established in common than in ad hoc categories. The
Type by Category Type interaction (min F' < I). Intru- current experiment attempts to show that instance-to-
sions did not differ during free recall between random concept associations are better established for common
category subjects and common category subjects than for ad hoc categories. Subjects received sets of
[F(I ,33) = 2.22, p > .10] or between random category exemplars to categorize. For example, ''What category
subjects and ad hoc category subjects (F < 1).4 do moth, bee, gnat, and ant all belong to?" ''What
category do coffee, perfume, leather, and skunk all
Discussion belong to?,,6 Each set contained exemplars from a single
Overall, common category subjects recalled more common category, exemplars from a single ad hoc cate-
words than did ad hoc category subjects both for free gory, or unrelated items. Half the subjects received a
and cued recall. Decomposing overall recall into cate- context vignette prior to each set that described a
gory access and exemplar recall yielded similar results: character trying to achieve a goal. The category instanti-
Common category subjects accessed more categories ated by the exemplars was usually instrumental to
than did ad hoc category subjects during free recall and achieving the goal; that is, the context primed the cate-
retrieved more exemplars per accessed category during gory. The remaining subjects received no such contexts.
both free and cued recall. Common category subjects All subjects tried to generate a category label for each
also produced more intrusions during free and cued set and rated how easy it was to do so.
recall than did ad hoc category subjects. All these find- If instance-to-concept associations are not well
ings are consistent with the conclusion that ad hoc established for ad hoc categories, then subjects without
categories are not as well established in memory as context should have difficulty labeling the ad hoc
common categories are. The concepts for common cate- category sets and should show much variability in the
gories and their concept-to-instance associations are labels they generate. This is because the ad hoc category
better established in memory than are those for ad hoc concepts should be difficult to activate, and subjects
categories." may activate a wide range of concepts in the process
There is an alternative explanation for the difference of trying to classify these sets. Difficulty should be
in accessibility between ad hoc and common categories. reduced, however, when these sets are preceded by
Namely, the category labels were longer for the ad hoc contexts that prime the category concepts. In contrast,
categories and, therefore, may have been more difficult the common categories should be easy to label even
to retrieve. But if retention is related to length of cate- without context, since their instance-to-category associa-
gory label, then this relation should hold within each tions are well established in memory. As each item in a
category type. The correlation between number of set is encoded, it should activate the same highly associ-
letters for each category label and probability of ated category concept. Consequently, subjects should
accessing the category was computed separately for the find it easy to categorize these sets and should show
ad hoc and common categories. The correlations were much consistency in the labels they generate. It would
222 BARSALOU

not be surprising if the common category sets were as Results


easy to categorize without context as with it. Ratings of ease and appropriateness. For each mea-
sure, one ANOVA was performed on the ad hoc cate-
Method gories and another on the common and random cate-
Design. Subjects attempted to categorize 14 item sets. Eight
were from ad hoc categories, three were from common cate- gories together. Planned comparisons contrasted means
gories, and three, as best as possible, were from no category. of interest between the two analyses. The results for the
For each set, subjects provided (1) a label categorizing all four ratings of ease and appropriateness were equivalent in
items (if one could be discovered), (2) a rating of how easy it effects. Therefore, only the tests for ease will be
was to find the label, and (3) a rating of how confident they were
that the label was appropriate.
reported, although the means for both ratings are shown
Half the subjects received a context vignette before each in Table 6. As recommended by Clark (1973) and Winer
item set, and half received no specified context. For the ad hoc (1971), subjects and categories were both treated as ran-
and common categories, each contex t provided a relevant setting dom effects in these analyses when appropriate. For the
for the category. The random categories bore no relation to their ad hoc categories, subjects found it easier to discover
contexts. Half the subjects received the ad hoc categories in one
random order, and half received them in the inverse order. For
a category label with context then without [F'(l ,12) =
all subjects, the common categories occurred at Positions I, 7, 16.02, p<.OI]. There was a Context by Categories
and 10, and the random categories occurred at Positions 2, 5, interaction [F( 7,140) = 6.52, p < .001] : Some ad hoc
and 11. These categories provided subjects with examples of the categories were more difficult to discover than others,
easiest and most difficult categories initially and intermittently but less so with context than without. This interaction
during the experiment; this allowed subjects to respond to the
ad hoc categories relative to the anchor points established by the occurs in all analyses for ad hoc categories reported here,
common and random categories. and its interpretation is the same in all cases.
Materials. Each subject received a booklet of instructions and Context had no effect on the ease of categorizing the
category materials. For the context condition, each page of common and random categories [F'(l ,9) = 1.00] , and
materials contained the context vignette, a blank line on which
to write a category label (if discovered), a column of randomly
there was no Context by Category Type interaction
ordered labels for the four exemplars, a response scale for indi- (F ' < 1). However, common categories were much easier
cating ease of category discovery, and a response scale for than random categories [F'(l,140) = 308.39, p < .001).
indicating how appropriate the label was for the four items. Although position in the list was not controlled for the
The two response scales were the integers from I to 7, 7 repre- common and random categories, varying this factor for
senting maximum ease and appropriateness, respectively. For the
no-context condition, the category materials did not contain the ad hoc categories did not have an effect or interact
the context vignettes, but they were otherwise identical to the with context. This suggests that position was not a fact-
materials for the context condition. tor for the common and random categories as well, and
The eight ad hoc categories were the same as those in Experi- that the context and category type results are valid for
ment 1. The common categories were "fruit," "birds," and the common and random categories.
"sports." The random categories were sets of items that intui-
tively did not constitute any category. Each context vignette With context, the ad hoc categories were as easily
described a character engaged in a goal-directed activity and discovered as the common categories [min F'(l ,37) =
primed the subsequent category. None of the vignettes contained 1.88, p > .101 and were more easily discovered than the
the category label for the respective item set. Table 1 contains random categories [minF'(I,31)=123.27, p<.OOI)
examples of the context and item sets. Without context, the common categories were easier to
Procedure and Subjects. Subjects in the context condition
were asked to read each vignette and to find, if they could, a discover than the ad hoc categories [min F'(l ,25) =
category to which all four items belonged. In the no-context 10.32, P < .01], and the ad hoc categories were easier
condition, subjects were instructed to study the four items and than the random categories [min F'(l ,30) = 6.16, P <
to find a category to which all belonged. All subjects were told .05).
that phrases, as well as single words, could serve as category
labels. They were also told that some item sets did not form a
Subjects' categorizations. Subjects generated labels
category, and that if no category was apparent, they should for the common categories 100% of the time. For the
write "0" on the blank line above the items. Once they had ad hoc categories, they did so 97% of the time with
generated a label or given up, they were to choose values on the context and 83% of the time without. For the ran-
ease and appropriateness scales. dom categories, they generated labels 14% of the time
The subjects, 24 undergraduates participating for either
course credit or pay, were randomly assigned, 6 each, to the with context and 31% of the time without. With so few
four context by order conditions. They worked through the labels, the random categories were not considered in
booklets at their own pace. the following analysis.

Table 6
Effects of Context and Category Type on Average Ratings of Category Discovery (Experiment 4)
Ease Appropriateness
Condition Random Ad Hoc Common Random Ad Hoc Common
Context 1.31 6.33 6.83 1.31 6.08 6.72
No Context 1.97 4.29 6.94 1.89 3.43 6.83
AD HOC CATEGORIES 223

How much did different subjects' labels for a given perceiving a category was positively correlated with how
item set denote the same category? For this analysis, well subjects agreed in labeling it. A correlated .94 with
the 12 (or sometimes fewer) category labels generated ease and .92 with appropriateness (both significant at
with context for each item set were typed in a ran- the Q = .001 level). These correlations show that
dom order on a page, and those generated without con- increased difficulty in discovering a category led to
text for the same item set were typed on another. The increased variability in subjects' categorizations.
resulting 22 pages were randomly ordered, half the
booklets having the inverse order of the others. Discussion
Six judges were asked to group all the labels on a page These data support the conclusion that instance-to-
that denoted the same category for as many categories as concept associations are better established in memory
were perceived. To be grouped, two or more labels had for common than for ad hoc categories. Without con-
to be very similar conceptually, but they did not neces- text, the ad hoc categories were difficult to identify, and
sarily have to share the same linguistic form. For each subjects were highly variable in the categories they dis-
category concept having two or more instantiations, covered. These category concepts only became obvious
the judge wrote the category concept and indicated the and agreed upon in relevant contexts that primed the
labels instantiating it. concepts.
A normalized measure of agreement, A, was com- Because ad hoc categories are so specialized, it may
puted for each judge's analysis of the labels for each be optimal that perceiving an entity does not activate all
item set in each context condition. A is defined as the ad hoc categories to which it belongs. Seeing a chair
[(number of category labels + I)-number of category and having categories such as "emergency firewood:' fits
concepts 1(number of category labels. "Number of cate- in the trunk of a car ," and "used to prop doors open"
gory concepts" equaled the number of concepts a judge come to mind would be highly distracting when these
thought were instantiated by one or more labels for an categories are irrelevant. Ad hoc categories should come
item set. A equaled 1 when all the labels were instantia- to mind only when primed by current goals. Such prim-
tions of one and only one category concept. A equaled ing does occur, as found in this experiment.
l(number of category labels, at the lower bound, when In contrast to ad hoc categories, context had no
no category concept had more than one instantiation." impact on ease of discovering common categories. The
The means from these analyses are shown in Table 7. concepts for these categories were as available without
For the ad hoc categories, agreement was greater with context as they were with context. This shows that
context than without (.90 and .59, respectively) instance-to-category associations are much better estab-
[F'(1,9) = 16.08, P < .01] ; this corresponds to slightly lished in memory for common than for ad hoc cate-
more than two concepts per set with context and gories. Interestingly, subjects with relevant contexts
slightly less than six concepts without, for a set of 12 were more variable in their categorizations of common
labels. There was also a Context by Categories interac- category item sets than were subjects without relevant
tion [F(7,28) = 7.44, P < .001] . contexts. It appears that the contexts caused subjects to
Surprisingly, for the common categories, there was be more specific in these categorizations and that peo-
more agreement without context than with context; this ple in general may often tailor common categories to
difference was marginally significant [F'(l ,6) = 5.28, current contexts. That is, categories like "fruit," "furni-
.10 > p > .05]. Subjects' labels with context were pri- ture ," and "clothing" may often be incorporated into
marily phrases relating the common category label, a ad hoc categories relevant to current purposes (e.g., fruit
single word, to the relevant character or context (e.g., for dessert, furniture to be moved, clothing in the laun-
John's favorite fruit, birds Mike saw). In contrast, sub- dry).
jects' labels without context were almost always the sin- A given entity can be cross-classified into an indef-
gle label for the category (e.g., fruit, birds). Subjects initely large number of categories. For example, "apple"
in the context condition, by being more specific, were can be cross-classified into "fruit," "things to take on a
more variable in their labelings. picnic," "things that could fall on your head," and so
The relation between category discovery and labeling. on. The data from the current experiment, in conjunc-
Categories that were difficult to label were also the ones tion with Barsalou's (1982) distinction between context-
that showed the least consistency in labeling. Ease of independent and context-dependent properties, suggest
the following account of cross-classification. During the
classification of an entity, categories with strong
Table 7
Effects of Context and Category Type on Agreement instance-to-category associations may be automatically
for Category Labeling (Experiment 4) activated. (e.g., "apple" may automatically activate
"fruit"). Alba, Chromiak, Hasher, and Attig (1980) and
Condition Ad Hoc Common
Ross and Barsalou (note I) provide further evidence that
Context .90 .89 such automatic classifications exist. In contrast, cate-
No Context .59 .98
- - _.. _-_.--_ .. gories weakly associated to an instance (e.g., "things to
Note-See text for description of agreement measure. take on a picnic" for "apple") may be activated only in
224 BARSALOU

contexts that require use of the category (e.g., going on inition I have been using. Even though they still violate
a picnic). These context-dependent categories are not correlational structure, their representations in memory
activated by the instance alone, but only by the con- are much more like those of common categories. Certain
junction of the instance and a particular context. In gen- ad hoc categories appear to make this transition. "Things
eral, classifications highly associated to an instance are to sell at a garage sale" may start out as ad hoc for some-
available across all contexts, whereas weakly associated one's first gargae sale but then become well established
classifications are only available in contexts that prime with subsequent ones. Similarly, someone taking up
them. camping may have the category of "things to take on a
camping trip" shift from being ad hoc to well estab-
GENERAL DISCUSSION lished.
A phenomenon discovered by Alba et a1. (1980) can
Categories other than the common taxonomic cate- be used to demonstrate this shift. These investigators
gories usually studied possess graded structure. More- presented subjects with instances of common categories
over, the ad hoc categories observed here had typicality ostensibly to learn for free recall. The number of
gradients as salient as those in common categories: instances presented per category ranged from three to
Ad hoc category instances varied as much in typicality as nine. Immediately following presentation, subjects were
common category instances, and subjects showed equal unexpectedly asked to estimate the number of instances
agreement when judging typicality for both category presented for each category (i.e., category frequency).
types. As suggested earlier, the same similarity compari- Across a wide range of instructions, list organizations,
son process appears to construct graded structure in and retrieval settings, subjects showed an unchanging
both common and ad hoc categories. Interestingly, this and excellent sensitivity for category frequency. Alba
process appears unaffected by how well established et al attribute this to automatized instance-to-concept
a category is in memory. associations: Every time an instance is encoded, it auto-
Although the ad hoc and common categories were matically activates its category concept. Sensitivity to
similarly structured, they differed in category represen- category frequency results from a count being kept at
tation. First, strong concept-to-instance associations for each category concept of how often it has been activated.
the common categories resulted in high consistency and Ross and Barsalou (Note 1) performed a similar
fast access during exemplar production and facilitated experiment with ad hoc categories. They initially found
the encoding and retrieval of relevant information pre- no sensitivity for category frequency, which they attrib-
sented for learning. Second, strong instance-to-concept ute to ad hoc categories not having strong instance-to-
associations in common categories resulted in highly concept associations. But if subjects repeatedly process
available category concepts that facilitated categoriza- these categories for a week before the surprise frequency
tion. And third, well established category concepts in test, strong instance-to-concept associations develop, and
common categories facilitated locating these categories equal sensitivity for ad hoc and common categories
during memory search. In contrast, lack of strong results. What were once ad hoc categories are no longer
concept-to-instance associations for ad hoc categories ad hoc.
resulted in less consistent and slower instance retrieval Similar to the potential for ad hoc categories to
during exemplar production and made it harder to become well established in memory is the necessity of
remember relevant information presented for learning. common categories being poorly established in early
Lack of strong instance-to-concept associations in ad hoc childhood. Horton (1982) demonstrated this by show-
categories resulted in less available category concepts, ing that children's poor performance with common cate-
thereby making categorization difficult. And the cate- gories results at least in part from a lack of well estab-
gory concepts for ad hoc categories were so poorly estab- lished memory structure. Children assessed as having
lished that they were no more accessible than the well established category representations did much bet-
representations of random word groups. Although ter on standard taxonomic tasks than children assessed
ad hoc and common categories may be constructed by a as having weak category representations.
shared process that bestows the same graded structure As discussed earlier, ad hoc and common categories
on each, other processes appear to result in the repre- differ in how well they reflect correlational structure.
sentations of common categories becoming better estab- Ross and Barsalou's (Note 1) category frequency results,
lished in memory. together with Horton's (1982) developmental find-
ings, demonstrate that how well a category reflects corre-
The Loss of Ad Hoc Status lational structure does not affect fundamental ways the
Some ad hoc categories may be processed so fre- category is processed. Both ad hoc and common cate-
quently that their category concepts, concept-to- gories can behave as categories that are either poorly
instance associations, and instance-to-concept associa- established or well established in memory. It does not
tions all become well established in memory. At this appear that the differences between ad hoc and common
point, these categories are no longer ad hoc by the def- categories in Experiments 2a, 3, and 4 regarding estab-
AD HOC CATEGORIES 225

lishment of category representation stem from how Constructing Category Concepts for Ad Hoc Categories
well the two reflect correlational structure. Instead, Perhaps one of the most difficult problems regarding
these differences most likely result from common ad hoc categories is explaining how goals make particular
categories' having received much more processing prior ad hoc categories relevant. In the previous section, I
to the experiments than ad hoc categories. Furthermore, assumed that the concepts for these categories contain
Experiments 1 and 2b demonstrate that reflecting cor- properties relevant to the goals these categories serve.
relational structure is not necessary for a category to But, given a goal, how can we predict which properties
exhibit graded structure. In general, the same similarity should be activated and associated to each other to form
comparison and memory processes appear to operate on the concept for an ad hoc category? The solution to this
categories that do and do not reflect correlational struc- problem will most likely need to be developed in the
ture. context of a theory of problem solving.
A more tractable problem is discovering how new
Determinants of Typicality concepts can be constructed to represent ad hoc cate-
As discussed earlier for common categories, how gories. One way is to retrieve well established concepts
similar an instance is to all other category instances and alter them by adding or deleting properties. In
(i.e., its family resemblance) determines its typicality. finding furniture for an office, "desk" and "lamp" may
The comparison-network model explains this relation- become "desk with a large surface" and "lamp that has
ship by assuming that a category concept is the aver- adjustable positions and provides bright light." Adding
age of all category instances and that an instance's properties was observed in Experiment 4 when subjects
typicality increases as it becomes more similar to the provided highly specific labels for common category
category concept. This model goes on to explain typi- item sets instead of using the one-word labels. Properties
cality in ad hoc categories in a similar manner; namely, can also be deleted from concepts when contrary to cur-
typicality depends on similarity to category concepts. rent goals. A dieter might order a "salad without dress-
Missing from this account, however, are assumptions ing," or someone who likes to work on furniture might
regarding the content of category concepts for ad hoc buy an "unupholstered sofa."
categories. Are they, too, averages of their category Novel concepts can also be constructed from prop-
instances? Barsalou (1981) provides evidence that they erty information in memory. Any property, X, can be
are not. In two studies, family resemblance bore no used to construct the category of "things possessing X."
relation whatsoever to typicality for categories that For example, the property of "flammable" could be
violate correlational structure. Instead, these cate- used to represent the category of "things that are flam-
gories are structured by dimensions relevant to the mable" for someone trying to prevent fires. Such concepts
goals the categories serve. For example, the dimensions may also contain sets of properties (e.g., things that are
of "calories" structures "things not to eat on a diet," flammable and near a heat source). Sets of properties
with the typicality of an instance increasing as its num- can be organized as conjunctions (e.g., expensive and
ber of calories increases. This dimension is important, unusual), as disjunctions (e.g., expensive or unusual), or
presumably because it is relevant to the goal the cate- as combinations of other forms.
gory serves, namely, losing weight. Once a new concept has been constructed, it may be
The comparison-network model accounts for this reconstructed on subsequent occasions if it continues to
finding as follows. The category concept for an ad hoc be relevant. Such reconstructions may cause the concept
category does not contain the average properties of its to become well established in memory. The associations
instances but, instead, only contains properties of the between its properties (if there are more than one) may
instances relevant to the goal the category serves. Since become stronger, and its overall accessibility may
only "edible" and "high in calories" are relevant to become higher.
losing weight, all other properties in "things not to eat on
a diet" are not included in the category concept. During Exemplar Production from Ad Hoc Categories
typicality judgments, the primary way instances can vary Once a concept has been constructed or retrieved,
from this concept in similarity is simply with regard to it can be used as a cue to retrieve category instances.
calories. Instances having few calories are not as similar For a category well established in memory, direct asso-
to the category concept as those high in calories and are ciations from its category concept to its category mem-
therefore less typical. Family resemblance does not bers are activated, thus activating the concepts for these
enter into typicality because the category concept does members. But for categories not well established in
not contain averages across all properties. All properties memory, there are no well established concept-to-
except those relevant to the goal are effectively weighted instance associations that serve this purpose. How, then,
to zero during the similarity comparisons. Although typ- do people retrieve exemplars from ad hoc categories?
icality derives from the same comparison process for One possibility is that they use a generate-test proce-
common and ad hoc categories, fundamentally different dure. The associative structure of related, well estab-
category concepts result in fundamentally different lished categories may be used to generate possible
forms of graded structure. instances of a poorly established category. As each item
226 BARSALOU

is retrieved, it is then tested for membership in the intelligence, understanding these processes may be cen-
poorly established category. To find instances of "res- tral to it.
taurants to watch a sunset in," instances from well
established restaurant categories could be retrieved (e. g., REFERENCE NOTE
local Indian restaurants, local seafood restaurants). Each 1. Ross, B. H., & Barsalou, L. W. Explorations into the nature
instance would then be checked for properties such as of the category frequency effect. Work in progress, University of
"has a western exposure," "has large windows," and Illinois, Champaign-Urbanna, and Emory University.
"has an unobstructed view." Instances having these prop-
erties become members of the category, and instances not REFERENCES
having them become members of its complement. ALBA, J. W., CHROMIAK, W., HASHER, L., & ATTIG, M. S.
As a category becomes frequently instantiated using a Automatic encoding of category size information, Journal of
generate-test procedure, direct associations should Experimental Psychology: Human Learning and Memory, 1980,
become well established between the category concept 6,370-378.
ANDERSON, J. R., & BOWER, G. H. Human associative memory.
and its instances. This change in category representa-
Washington, D. C: Winston, 1973.
tion would eventually make the generate-test proce- BARSALOU, L. W. Determinants of graded structure in categories.
dure unnecessary, since the more efficient lookup pro- Unpublished doctoral dissertation, Stanford University, 1981.
cedure could now operate. BARSALOU, L. W. Context-independent and context-dependent
information in concepts. Memory & Cognition, 1982, 10, 82-93.
BATTIG, W. F., & MONTAGUE, W. E. Category norms for verbal
Cross-Classification into Ad Hoc Categories items in 56 categories: A replication and extension of the Con-
Rosch, Mervis, Gray, Johnson, and Boyes-Braern necticut category norms. Journal of Experimental Psychology,
(1976) argue that people prefer to classify entities 1969,80, (Monograph Supplement 3, PI. 2).
initially with basic category names. For example, people BOUSFIELD, W. A. The occurrence of clustering in the recall of
randomly arranged associates. Journal of General Psychology,
prefer to call an inanimate object with four legs, a seat,
1953,49,229-240.
and a back in a particular configuration a "chair" as BOUSFIELD, W. A., & COHEN, B. H. The effects of reinforcement
opposed to an "office chair" or "furniture." Given this on the occurrence of clustering in the recall of randomly arranged
initial classification, however, there are numerous ways associates. Journal ofPsychology, 1953,36,67-81.
an entity can be cross-classified during subsequent classi- BOWER, G. H., CLARK, M. C., LESGOLD, A. M., & WINZENZ, D.
Hierarchical retrieval schemes in recall of categorized word lists.
fications that serve particular goals. For example, a chair
Journal of Verbal Learning and Verbal Behavior, 1969, 8,
could subsequently be classified into "things that can 323-343.
be used for emergency firewood," "things that can be CLARK, H. H. The language-as-fixed-effect-fallacy: A critique of
stood on," "gifts," and so on, depending on the current language statistics in psychological research. Journal of Verbal
goal of the perceiver. An outline of a cross-classification Learning and Verbal Behavior, 1973,12,335-359.
COFER, C. N. Does conceptual organization influence the amount
model was sketched in the discussion of Experiment 4. retained in immediate free recall. InB. Kleinmuntz (Ed.), Con-
How well someone can generate secondary cross- cepts and the structure of memory. New York: Wiley, 1967.
classifications appears to vary substantially. In fact, COLLINS, A. M., & LOFTUS, E. F. A spreading-activation theory
the ability to do this well has often been considered of semantic processing. Psychological Review, 1975,81,407-428.
CRONBACH, L. J. Essentials of psychological testing. New York:
a sign of creative ability. In the Unusual Uses Test Harper & Row, 1970.
(Cronbach, 1970), subjects are given the name of an GUILFORD, J. P., & FRUCHTER, B. Fundamental statistics in psy-
entity and asked to generate as many uses of it as they chology and education. New York: McGraw-Hill, 1973.
can think of, each of which can be considered a cross- HALL, J. F. Learning as a function of word-frequency. American
classification. The more cross-classifications a person Journal ofPsychology, 1954,67, 138-140.
HAYES-RoTH, B. Evolution of cognitive structures and processes.
generates, and the more novel these classifications are, Psychological Review, 1977,84,260-278.
the more creative the person is assessed as being. HORTON, M. Category familiarity and taxonomic organization in
In general, the construction and use of ad hoc cate- young children. Unpublished doctoral dissertation, Stanford
gories appear to reflect creative aspects of human intelli- University, 1982.
KUCERA, H., & FRANCIS, W. N. Computational analysis of
gence. Similar to the ability to cross-classify, the abilities present-day English. Providence, R. I: Brown University Press,
to construct new concepts instrumental to achieving 1967.
goals and to retrieve instances of these concepts without LOFTUS, E. F. Spreading activation within semantic categories:
direct concept-to-instance associations are creative pro- Comments on Rosch's "Cognitive representations of semantic
cesses. All three enable the construction of new repre- categories." Journal of Experimental Psychology: General,
1975,104,234-240.
sentations, each representation reflecting a new way of LORCH, R. F., JR. Priming and search processes in semantic mem-
organizing the environment. Perceiving these new organi- ory: A test of three models of spreading activation. Journal of
zations may be necessary to achieving new goals or to Verbal Learning and Verval Behavior, 1982,11,468-492
approaching old ones in novel ways. Once an ad hoc tax- MANDLER, G. Organization and memory. In K. W. Spence &
onomy has been constructed, further use may cause it to J. T. Spence (Eds.), The psychology of learning and motivation
(Vol. 1). New York: Academic Press, 1967.
become well established in memory. Understanding the MANDLER, G., PEARLSTONE, Z., & KOOPMANS, H. S. Effects of
construction, use, and establishment of ad hoc categories organization and semantic similarity on recall and recognition.
may not only turn out to be informative about human Journal of Learning and Verbal Behavior, 1969, I, 410-423.
AD HOC CATEGORIES 227

MCCLOSKEY, M., & GLUCKSBERG, S. Natural categories: Well per item set, represent the probability of a subject's circling an
defined or fuzzy sets? Memory & Cognition, 1978,6,462-472. item while guessing. Then if n is the number of subjects in the
MCCLOSKEY, M., & GLUCKSBERG, S. Decision processes in veri- experiment, pn is the number of subjects circling a given item
fying category membership statements: Implications for models and (1 - p)n is the number of subjects not circling the item.
of semantic memory. Cognitive Psychology, 1979,11, 1-37. Agreement for the item, as defined in the text, is thenl(+l)pn+
MERVIS, C. B., CATLIN, J., & ROSCH, E. Relationships among (-1)(1 - p)nl/n, which is 2p - 1. Averaging across items and
goodness-of-example, category norms, and word frequency. item sets results in an overall agreement score for the experiment
Bulletin of the Psychonomic Society, 1976, 7, 283-294. of 2p - 1. Given that p in Experiment 1 was .52 (i.e., the aver-
MERVIS, C. B., & RoSCH, E. Categorization of natural objects. age number of items circled per set, 3.15, divided by the number
Annual Review ofPsychology, 1981,32,89-115. of items in a set, 6), then overall agreement should be .04 if sub-
ORTONY, A. Beyond literal similarity. Psychological Review, 1979, jects were guessing.
86, 161-180. 2. It should be noted that the unclear cases observed in these
PUFF, C. R. Role of clustering in free recall. Journal of Experi- categories were unclear in the sense that some subjects thought
mental Psychology, 1970,86,384-386. these items were category members and other subjects did not.
RATCLIFF, R., & McKoON, G. Does activation really spread? Psy- As noted by McCloskey and Glucksberg (1978), such cases are
chological Review, 1981,88,454-462. not necessarily unclear. They could falsely appear to be unclear
RIPS, L. J., SHOBEN, E. J., & SMITH, E. E. Semantic distance because some subjects' definitions of a category include these
and the verification of semantic relations. Journal of Verbal items as category members, whereas other subjects' slightly dif-
Learning and Verbal Behavior, 1973, 12, 1-20. ferent definitions do not. Most important, classifying these
ROSCH, E. H. On the internal structure of perceptual and semantic instances may be completely clear (as opposed to unclear) for all
categories. In T. E. Moore (Ed.), Cognitive development and subjects. Given the presence of typicality gradients in these cate-
the acquisition of language. New York: Academic Press, 1973. gories, however, it is likely that the unclear cases in this experi-
ROSCH, E. H. Cognitive representations of semantic categories. ment really were items subjects were uncertain about. If sub-
Journal of Experimental Psychology: General, 1975, 104, jects show decreasing confidence in the membership of clear
192-233.(a) category members, which they did, it is reasonable to believe
ROSCH, E. H. The nature of mental codes for color categories. their confidence would decrease to the point of being uncer-
Journal of Experimental Psychology: Human Perception and tain about the membership of unclear cases.
Performance, 1975,1,303-322. (b) 3. That subjects ably discriminate ad hoc categories is corro-
ROSCH, E. H. Reply to Loftus. Journal of Experimental Psycho- borated by tile following demonstration. Six subjects generated
logy: General, 1975,104,241-243. (c) four exemplars to each of the eight ad hoc categories used in
ROSCH, E. H., & MERVIS, C. B. Family resemblances: Studies in Experiment 1. Four judges examined each of the 192 exemplars
the internal structure of categories. Cognitive Psychology, 1975, generated as to whether they were valid members of their respec-
7,573·605. tive categories. Three judges accepted all 192 exemplars as
RoSCH, E. H., MERVIS, C. B., GRAY, W. D., JOHNSON, D. M., & valid, and one judge accepted all but 4. Thus there appears to be
BOYEs-BRAEM, P. Basic objects in natural categories. Cog- substantial agreement on membership in ad hoc categories.
nitive Psychology, 1976,8,382-439. 4. It was impossible to compute item averages for the random
ROSCH, E. H., SIMPSON, C., & MILLER, R. S. Structural bases of category intrusion data since it was not clear to which category a
typicality effects. Journal of Experimental Psychology: Human given intrusion belonged. Therefore, the one-way ANOVA for
Perception and Performance, 1976,2,491-502. the intrusion frequencies was performed only across subjects.
SMITH, E. E., & MEDIN, D. Categories and concepts. Cam- 5. Since number of correct exemplars retrieved and number
bridge, Mass: Harvard University Press, 1981. of intrusions were both higher for common than for ad hoc cate-
SMITH, E. E., SHOBEN, E. J., & RIPS, L. J. Structure and pro- gories, one could argue that common categories are not superior
cess in semantic memory: A featural model for semantic deci- mnemonic devices. However, the total increase during free recall
sions. Psychological Review, 1974,81,214-241. in the number of correct exemplars from ad hoc to common
TULVING, E. Subjective organization in free recall of "unrelated" categories was 7.68, whereas the total increase in the number of
words. Psychological Review, 1962,69,344-354. intrusions was only 1.50. For cued recall, the increase in cor-
TULVINO, E., & PEARLSTONE, Z. Availability versus accessibility rect exemplars was 6.72, whereas the increase in intrusions was
of information in memory for words. Journal of Verbal Learn- 2.06. The much larger increases for correct exemplars indicate
ing and Verbal Behavior, 1966,5,381-391. that common categories do indeed make superior mnemonic
TvERSKY, A. Features of similarity. Psychological Review, 1977, devices.
84,327-352. 6. The categories are "insects" and "things that have a
WINER, B. J. Statistical principles in experimental design. New smell," respectively.
York: McGraw-Hill, 1971. 7. The minimum increased as the number of labels for a set
ZADEH, L. A. Fuzzy sets. Information and Control, 1965, 8, decreased. However, this worked against the hypothesis that
338-353. there would be a context effect, since the no-context sets usually
had fewer labels than the context sets.
NOTES
(Manuscript received for publication March 26, 1982;
1. Let p, defined as the average proportion of items circled revision accepted February 7,1983.)

You might also like