ArticlePDF Available

Converging operations on a basic level in event taxonomies

Authors:
Article

Converging operations on a basic level in event taxonomies

Abstract and Figures

Research on object concepts has identified one level of abstraction as "basic" in cognition and communication. We investigated whether concepts for routine social events have a basic level by replicating the converging operations used to investigate object concepts. In Experiment 1, subjects were presented with event names from a taxonomy and were asked to list the actions comprising the event. Many more actions were listed at the middle than at the highest taxonomic level, without a further increase at the most specific level, paralleling the pattern of superordinate-, basic-, and subordinate-level object concepts. From these action lists, brief stories were composed for each event. In Experiment 2, subjects made pairwise similarity judgments on the stories. The mean similarity of events increased with specificity, as expected. But differentiation of categories (within-category similarity compared to between-category similarity) was highest for super-ordinates, contrary to results with object categories. In Experiment 3, subjects were fastest in recognizing actions as belonging to events named at the basic level. In Experiment 4, subjects predominantly chose basic-level terms to name stories. We conclude that event taxonomies do show basic-level structure, albeit a less sharply defined and less stable structure than in object taxonomies. The benefits and hazards of extending models of object concepts to other entities, such as social events, are discussed.
Content may be subject to copyright.
Memory
&
Cognition
1990,
18 (4),
407-418
Converging operations on a basic level
in event taxonomies
MICHAEL W. MORRIS
University
of
Michigan,
Ann
Arbor, Michigan
and
GREGORY
1.
MURPHY
Brown University, Providence, Rhode Island
Research on object concepts
has
identified one level of abstraction as "basic" in cognition and
communication. We investigated whether concepts for routine social events have a basic level
by replicating
the
converging operations used to investigate object concepts. In Experiment I,
subjects were presented with event names from a taxonomy and were asked to list
the
actions
comprising
the
event. Many more actions were listed at the middle
than
at
the
highest taxonomic
level, without afurther increase at the most specificlevel, paralleling the pattern ofsuperordinate-,
basic-,
and
subordinate-level object concepts. From these action lists, brief stories were composed
for each event. In Experiment 2, subjects made pairwise similarity judgments on
the
stories. The
mean similarity of events increased with specificity, as expected. But differentiation ofcategories
(within-category similarity compared to between-category similarity) was highest for super-
ordinates, contrary to results with object categories. In Experiment 3, subjects were fastest in
recognizing actions as belonging to events named
at
the
basic level. In Experiment 4, subjects
predominantly chose basic-level terms to name stories. We conclude
that
event taxonomies do
show basic-level structure, albeit a less sharply defined and less stable structure
than
in object
taxonomies. The benefits and hazards of extending models of object concepts to other entities,
such as social events, are discussed.
The question,
"How
shall a thing be called?" led Roger
Brown (1958) to observe that out of the myriad words
that can describe an object, one level of generality stands
out as preferred. We call what we sit on a chair as op-
posed to a desk chair or a piece
of
furniture. The story
tells of the apple that fell on Newton's head, not the
granny smith or the piece of fruit. Brown concluded that
we name objects at the level most useful in most contexts,
using the name that generally distinguishes the referent
from other objects.
Eleanor Rosch and her colleagues (Rosch, 1978; Rosch,
Mervis, Gray, Johnson, &Boyes-Braem, 1976) provided
a more comprehensive answer with their proposal that ob-
jects are encoded in memory primarily at a medium level
of abstraction, which they called the basic level of con-
cepts, although more general superordinate and more
specific subordinate concepts also play roles in cognition.
Rosch et al. (1976) found that concepts at the basic level
are optimally informative and show a preferred status in
many cognitive processes, such as recognition and nam-
This research was supported by NIMH Grant MH-41704 and NSF
Grant BNS 83-15145 and was completed while the first author was at
Brown University. We thank Jean Liittschwagerfor her invaluablehelp
in running the experiments, and Ed Smith and the reviewers for their
helpfulcommentson the manuscript. Correspondence may be addressed
to Michael W. Morris, Department of Social Psychology, University
of Michigan, 340 Thompson St., Ann Arbor, MI48106.
ing. In this paper, we investigate this principle of category
structure beyond the domain of objects. Our experiments
were designed to investigate whether concepts for events
also have a basic level with similar processing advantages.
The basic level is a compromise. On the one hand, con-
cepts should be specific and informative, so that know-
ing what category a thing belongs to will reveal many at-
tributes. On the other hand, concepts should be general
and inclusive, in order to minimize the total number of
different categories a person needs. Basic-level concepts
resolve this tradeoff by optimally combining distinctive-
ness and informativeness (Murphy &Brownell, 1985;
Rosch et al., 1976). For example, consider the basic con-
cept table. Tables are distinct from other types of furni-
ture (they are not very similar to chairs, beds, lamps,
bureaus, etc.), and knowing that something is a table gives
important information about it (its shape, function, con-
struction, etc.). The subordinate coffee table is much less
distinctive, since coffee tables are very similar to other
kinds of tables. The superordinate furniture is much less
informative than the basic concept; knowing that some-
thing is a piece of furniture does not specify its shape,
parts, function, or materials to any degree. Thus, the only
level that is both distinctive and informative is the basic
level.
Rosch et al. (1976) had two goals in their studies of
the basic level. The first was to discover whether there
is one level of abstraction that maximizes the combina-
407 Copyright 1990 Psychonomic Society, Inc.
408 MORRIS AND MURPHY
tion of inclusiveness and informativeness. Their second
goal was to demonstrate that basic concepts have a
processing advantage (i.e., are used faster or more ac-
curately) over other concepts in a variety of tasks.
Toward their first goal, Rosch et al. (1976) performed
a number of feature-listing experiments. In their Experi-
ment 1, subjects listed the attributes that most objects in
a category have in common. In their Experiment 2, they
collected the actions associated with objects in a category .
The same pattern held in both experiments: subjects listed
few features for abstract taxonomic terms, many features
for the medium-level terms, and few additional features
for the most specific terms. Tversky and Hemenway
(1984) found qualitative as well as quantitative differ-
ences: distinctive features listed at the basic level are parts,
whereas superordinate features are mainly functional, and
subordinate features are mainly perceptual. In succeed-
ing experiments, Rosch et al. fulfilled their second goal,
finding that basic concepts were the most useful in a va-
riety of perceptual and linguistic tasks (see below).
Since Rosch et al. (1976) outlined the basic-level the-
ory and found evidence for it in the domain of objects,
others have extended it to domains farther and farther re-
moved from natural objects, including artificial stimuli
(Mervis &Crisafi, 1982; Murphy &Smith, 1982), com-
puter programming (e.g., sorting algorithms; Adelson,
1983), environmental scenes (e.g., beaches; Tversky &
Hemenway, 1983), psychological situations (e.g., work;
Cantor, Mischel, &Schwartz, 1982), and personality con-
cepts (e.g., comic jokers; Cantor &Mischel, 1979).
Basic-Level Concepts for Events
Events can be named at various levels of abstraction,
but as with objects, one level is generally preferred. For
example, an acquaintance describing her plans for the day
might report events like "taking ashower," "having
breakfast," and "playing tennis."
If
she reported
"per-
forming hygienic activities," "eating a meal," and "play-
ing a
sport,"
she would sound evasive. Alternatively,
if she reported "taking a cold shower," "having
Wheaties," and "playing Canadian doubles," one might
wonder what she was trying to emphasize with this extra
detail (Cruse, 1977). Such intuitions suggest that there
may be a basic level of event concepts.
Rosch (1978) was the first to suggest that the basic-level
theory might be extended to concepts for events. She con-
ducted a pilot study in which students were asked to list
the events of their day. They also listed events for smaller
and larger time units than days. The subjects reported
events at a consistent level of abstraction. They described
activities in terms of highly predictable sequences of ac-
tions, such as "feeding the
cat,"
and not in more general
terms, such as
"all
the morning chores," or more specific
terms, such as "opening the cat food." Although sug-
gestive, these results are inconclusive without the set of
converging operations necessary to define a basic level
(see Rosch et al., 1976). In particular, it is necessary to
establish a consistent taxonomy and then compare event
concepts from its different levels of abstraction.
The event concepts that Rosch tentatively called basic
seem similar to the scripts for routine events, such as go-
ing to a restaurant, proposed by Schank and Abelson
(1977). The research of Schank and his colleagues has
primarily concerned the internal structure of scripts rather
than the organization of event knowledge, and most of
the psychological literature has followed this lead (Ab-
bott, Black, &Smith, 1985; Bower, Black, &Turner,
1979). That is, rather than studying how scripts are related
to each other and to more general event representations
(i.e., studying taxonomies), they have investigated how
the parts of individual scripts are related and used (see
Barsalou &Sewell, 1985, for a discussion). Thus,
although our project is related to research on script
memory, it focuses on issues these researchers have not
addressed.
One previous researcher, Rifkin (1985), systematically
compared the levels of event taxonomies. Rifkin elicited
features of event categories and found an apparent basic
level. For compelling evidence, however, a series of con-
verging operations, as in Rosch et al. 's (1976) original
work, is required. To provide this evidence, we replicated
experiments on object concepts using an extended ver-
sion of Rifkin's taxonomies.
In Rosch et
al.'s
(1976) Experiment 2, subjects listed
the motor movements associated with an object; subjects
in our first experiment listed the actions that make up
events. In this way, the behavioral informativeness of
event terms can be measured at all three levels of the tax-
onomies. This experiment will also reveal the configura-
tion of parts in event concepts, since actions are parts of
events. Tversky and Hemenway (1984) showed that parts
are the distinctive features of basic-level object concepts,
and in Experiment 1 we investigated whether the same
holds for events. In Experiment 2, we compared between-
category and within-category similarity at each level of
event concept (Mervis &Crisafi, 1982) as an additional
measure of conceptual structure. Once the informative-
ness of event concepts is understood, we can then pro-
ceed to test for processing advantages. In Experiments
3 and 4, we investigated naming and categorization
processes.
EXPERIMENT 1
Much of the information associated with objects or
events is behavioral information. In Rosch et al. 's (1976)
Experiment
2-behavioral
measurement of category
structure-subjects listed the motor movements associated
with objects. In the current experiment, subjects listed the
actions that make up events. We measured the behavioral
informativeness of the three levels of event terms in Rif-
kin's (1985) taxonomies.
Method
Subjects. Thirty-five Brown University undergraduates were used
as subjects. Most volunteered their time; a few were paid for their
participation.
Materials. The stimuli were 45 taxonomically related event terms
generated by Rifkin's (1985) subjects. There were nine general (i.e.,
superordinate) terms, with two terms at a medium level
of
detail
for each general term and one specific term for each of the medium
ones (see Rifkin, 1985, for details on how the categories were de-
rived). This category structure has one drawback, namely that each
middle-level event category
(e.g.,
wash hands) has only one sub-
category
(e.g.,
wash hands quickly). Since some measures
of
category structure require two or more categories at each level, we
used
Rifkin's
raw data (published in his Appendix B) to select a
second specific event for every middle-level category. Out
of
the
five
or
so candidates for each subordinate, we chose the category
that was most distinctive from the one that Rifkin chose. The full
set is
shown
in Table I.
The
stimuli
were
divided into seven sets
of
nine items each: one
set
of
general terms, two sets
of
medium-level terms, and four sets
of
specific terms.
Each
set contained one item from each
of
the
nine taxonomies, and 5subjects responded to each set. Each sub-
ject
received abooklet with instructions on the
cover
and one con-
cept
term
at the top
of
each following page. The test booklets were
assembled with the nine pages
of
items in a different random
order
for each subject.
Procedure.
The
subjects received test booklets and read the in-
structions. They were then allowed 2min to respond to each event
term. The instructions called event
terms'
'activities," and the sub-
jects were asked to list the parts
of
each activity. The parts
of
an
activity were described as the
"actions"
common to all examples
of
that activity. The subjects were instructed not to write prose
descriptions, but rather lists
of
short phrases describing actions,
EVENT CONCEPTS 409
and to avoid responding in terms
of
personal associations. The in-
structions contained two examples of activities and their parts. De-
pending on the level
of
events that a subject received, he or she
saw the examples
"going
out to
eat"
and
"vacation"
(for those
who received general terms),
"restaurant"
and
"camping
vaca-
tion"
(middle-level terms), or
"fast
food
restaurant"
and
"family
camping vacation" (low-level terms). The example parts listed were
the same across levels.
Judgeamending. A
judge
(one
of
the authors) tallied all actions
listed for each concept by at least 2 out
of
the 5 subjects who saw
it. This criterion was similar to that used by Rosch et at. (1976).
Making these tallies required equating similar actions listed by differ-
ent subjects
(e.g.,
"pay"
and
"pay
for the
meal").
To check the
reliability
of
this process, asecond
judge
(the other author) tallied
the actions for five
of
the nine event taxonomies (the two highest,
the two lowest, and the median taxonomy in number
of
actions tal-
lied). There was high agreement
(r
=.87) between the action tal-
lies
of
the two judges. Next, a new group
of
judges
reviewed the
action lists to amend our lists in the manner
of
Rosch et at. and
Tversky and
Hemenway's
(1983)
"judge-amended
tallies."
These
previous researchers amended the lists to minimize artifaetual results
stemming from the pragmatics
of
the task. One possible artifact
sterns from subjects' following Grice's (1975) conversational maxim
of
relevance: Some actions listed for superordinate and basic-level
events may be parts of the subordinate event as well, but they might
have seemed too obvious for the subjects to mention. For example,
"get
a
weapon"
was listed
for"
Assault" but not
for"
Assault with
Superordinate
Table 1
Rifkin's (1985) Event Taxonomies
Basic Subordinate
Meal
Entertainment
School Activities
Sports
Hygiene
Shopping
Crime
Transportation
Housework
Breakfast
Dinner
Movies
Theatre
Classes
Tests
Football
Hockey
Wash Hands
Shower
Food Shopping
Furniture Shopping
Murder
Assault
Train Travel
Bus Travel
Vacuuming
Sweeping
Quick Breakfast
Going Out for Pancakes
Family Dinner
Banquet Dinner
Horror Movies
Western Movies
Comedy at the Theatre
Drama at the Theatre
English Class
Math Class
Essay Tests
Multiple-Choice Exams
Touch Football
Tackle Football
Professional Hockey
Olympic Hockey
Wash Hands Quickly
Rinsing Hands
Lockerroom Shower
Quick Shower
Shopping at a Butcher Shop
Shopping for Fruit
Shopping for Bedroom Furniture
Shopping for a Dining Room Set
Homocide
Patricide
Assault with a Deadly Weapon
Mugging
Subway Travel
Amtrak to New York
Travel by City Bus
Travel by School Bus
Vacuuming the Carpet
Vacuuming the Drapes
Sweeping the Floor
Sweeping a Rug
410 MORRIS
AND
MURPHY
EXPERIMENT
2
Table 2
Number
of Judge-Amended Actions Listed in Experiment 1
Level of Abstraction
Feature listing is a traditional method for describing the
information in concepts. However, the sheer number
of
features listed for a concept cannot be taken as an exact
measure
of
the information carried by that concept. One
obvious problem with this operationalization is that there
are known qualitative differences between the features
listed at various levels
of
abstraction: functional features
at the superordinate level, parts at the basic level, and
perceptual features at the subordinate level (Tversky &
Hemenway, 1984). Whether five functional features carry
less information than six perceptual features is difficult
to say. An alternative measure
of
information in a con-
cept is the similarity
of
the category's members. Similarity
the judges decided that an action was generally true
of
the basic-level event but not
of
the subordinate. Rosch
et al. (1976) used a forced inheritance rule in their tax-
onomies, such that any feature true
of
one category was
assumed to be true
of
all lower level categories. Although
feature inheritance is a property
of
all true taxonomies,
when features and categories are not very typical this
property does not hold in people's category representa-
tions.
For
example, although the subjects agreed that
washing
one's
hands normally includes using soap, they
also judged that washing
one's
hands quickly (a subor-
dinate) does not require soap. Lack
of
feature inheritance
is found in object categories as well (Hampton, 1987).
We decided to accept the
judges'
decisions without forc-
ing them into a taxonomic pattern
of
inheritance.
The number
of
features in both Rifkin's (1985) attribute-
listing task and our action-listing task increases dramati-
cally with increasing specificity to the basic level but not
thereafter. One might hesitate to accept these experiments
as independent, converging operations, since they are very
similar. This objection can be answered by Rifkin's report
that less than 14 %
of
his subjects' features were classifi-
able as actions. So the results from Experiment 1 con-
verge with Rifkin's to reveal basic-level structure in event
concepts.
15.0
16.0
14.5
12.0
9.2
16.5
4.5
12.0
10.8
12.3
14.2
16.0
12.8
11.5
10.2
16.2
5.0
11.8
9.8
11.9
10.5
2.0
3.5
4.8
0.5
11.8
1.8
4.0
2.5
4.6
Superordinate Basic Subordinate
Concept
M
Meal
Entertainment
School Activities
Sports
Hygiene
Shopping
Crime
Transportation
Housework
a Deadly
Weapon,"
probably because it seemed already implied
by the event name. Thus, the unamended results might have a bias
that inflates the size
of
the basic-level advantage. Conversely, the
items listed at the superordinate level could incorrectly include event
instances instead of event parts, as Rifkin (1985) found. This would
cause a bias that would deflate the basic-level advantage. Judge
amending removed such biases.
Three naive graduate students and an author acted as judges. They
were given 36 rating forms, each with one of the subordinate terms
and its basic and superordinate category listed across the top of the
page. All the actions given in the feature-listing task for the
subordinate-level name and its dominant superordinate and basic
terms were randomly listed down the left side of the page. The judges
were instructed to decide whether each action was part
of
each ac-
tivity. The instructions included Tversky and Hemenway's (1984)
standard for the goodness of object parts, extending it to event parts:
For an action to be a true part of an event, the judges were told,
it must be perceptually discriminable, functionally distinct, and more
or less obligatory to the activity. The judges also indicated when
two of the listed actions were redundant; that is, when they referred
to the same part of the event. The judges took the test booklets home
with them and were allowed 2 days to complete them. Ratings were
then compiled, with agreement of 3 out of the 4 judges as the crite-
rion for including a part. The judges convened to resolve ratings
of the items on which opinion was evenly divided. Items eliminated
were either instances rather than parts (as swimmingis to sports)
or redundant actions. The final score for each category was its mean
across all of the score sheets it appeared on.
Results
and
Discussion
On average, the subjects listed
6.1,
8.0, and 8.4 parts
for the high-, middle-, and low-level categories, respec-
tively. Although these results suggest abasic-level struc-
ture, such unamended means could be influenced by the
biases mentioned above, so we will only analyze in de-
tail the judge-amended means. We expected that the basic
level would be the most general level in event taxonomies
at which many parts are shared, and Table 2reveals that
this prediction was fulfilled. The three taxonomic levels
received different numbers
of
features [F(2,24) =12.13,
P<.001].
For
comparison with Rifkin's (1985) data,
we first analyzed only the categories he studied. The
middle-level event terms elicited significantly more parts
than did the superordinates [t(8) =8.33, p<.001]. The
number
of
parts listed for middle-level event terms and
for subordinate-level event names were nearly identical
[t(8) =0.13].
When
the data for all 36 subordinates
are included, the pattern is identical: the basic-level lists
were
significantly longer than the superordinate-level
lists [t(8) =
4.17,
p<.001], and the subordinate-level
lists were not significantly longer than the basic-level lists
[t(8) =0.19]. In other words, as event names increase
in generality from the subordinate level to the basic level,
there is no loss in part information. But a further increase,
to the more abstract superordinate terms, does create such
a loss. Thus, these results agree with those
of
Rifkin in
establishing a basic level
of
events.
One might be puzzled about the cases in Table 2 in
which there are actually fewer actions listed for
subordinate- than for basic-level events. In these cases,
EVENT CONCEPTS 411
Table 3
Mean
Rated
Similarity of Event Pairs versus Object Pairs
1.18
1.88
1.97
3.276.91
6.41 4.85
6.58 5.15
7.70
8.12
8.01
Relationship
Same Same Same
Subordinate Basic Superordinate Unrelated
Stimuli
Event Stories
Event Names
Pictures of
Natural Objects
Pictures of
Artificial Objects 7.73 6.91 2.85 1.38
Note-
The event data are from Experiment 2, and the object data
presented for comparison are from Mervis and Crisafi (1982). Rating
was done on a 1-9 scale.
Ninety story pairs were assembled: 36 same-subordinate pairs,
which were simply the two stories derived from each subordinate;
18same-basic pairs, which were stories from the samebasic category
but different subordinates; 18same-superordinate pairs, which were
stories from the same superordinate but different basic categories;
and 18unrelated pairs, which were stories from different superor-
dinates. The particular pairings were determined randomly whenever
possible. Each subordinate in the taxonomy appeared once in each
condition (same-subordinate, same-basic, etc.). Each pair of
paragraph-long stories was mounted on a large index card, and the
cards were randomly numbered
1-90.
Subjects. The subjects were 16 Brown University undergradu-
ates, who were paid for their participation.
Procedure.
Each subject received
half
of
the 90 story pairs, in
one
of
these orders:
1-45,
45-1,
46-90,
or
90-46.
The subjects
judged the similarity
of
the two stories using a 1(not at all similar)
to 9(extremely similar) scale. The instructions emphasized that the
subjects should read the stories thoroughly. The scale and instruc-
tions were the same as those used by Mervis and Crisafi (1982),
allowing us to compare our results to theirs.
Results
The mean similarity of events of all four kinds was com-
pared to Mervis and Crisafi's (1982) results for objects,
and is presented in Table 3. At a gross level, the patterns
are the same: similarity increased with increasing tax-
onomic proximity [F(3,45) =237.82, p<.001]. This
result is not very significant, however, since it shows only
that the more specificcategories were more similar, which
was never in doubt. At a finer level, the patterns diverge.
Superordinate event concepts seem coherent-their mem-
bers are similar relative to the unrelated concepts. This
difference became clearer when differentiation scores
were calculated following Mervis and Crisafi's method.
In that procedure, the similarity of objects within a
category has the similarity of objects between categories
at that level subtracted from it, providing a measure
of within- versus between-category similarity at each
category level. For example, the basic-level differentia-
tion score is the mean basic similarity minus the mean
superordinate similarity. Instead of finding the basic level
to be the most differentiated, as Mervis and Crisafi found
with objects (see Table 3), our results show the greatest
differentiation at the superordinate level. The scores were
2.97, 1.56, and 1.71 for superordinate, basic, and subor-
dinate categories, respectively. That is, superordinates
showed the greatest difference between within-category
The two distinct stories for each subordinate were derived by varying
the names
of
characters and places and by adding aunique action
to each. These additional actions were
drawn
from the raw data
in Experiment 1. They were actions listed for which there was in-
sufficient consensus to be included in the final action list. This ad-
dition was necessary to prevent the subordinate stories from being
completely identical.
Jean wanted to see a horror movie. She searched through the listings
in the paper for a good film, then picked a show time and called some
friends to go with her. Jean went to the theater and bought a ticket.
She bought somefood, and then gave her ticket to the usher as she en-
tered the movie room. She picked a seat and sat down. Jean watched
the preview andthe show, screaming during the scary parts. When it
was over, she left.
Method
Stimuli.
Two
brief
stories were composed corresponding to each
of
the 36 subordinate-level event terms listed by the subjects in Ex-
periment 1.
The
72 stories were simple accounts
of
acharacter do-
ing the event. Each action in the judge-amended lists (from Ex-
periment
1) for a subordinate
term
became aclause in the story.
The
subordinate
event
term
itself
occurred
in the first sentence
of
the story, since the subjects had listed actions in that context. For
example, astory constructed for the subordinate horror movies
follows:
Experiment 2A
Judging Event Stories
Mervis and Crisafi (1982) examined natural and artifi-
cial object concepts by comparing the similarity of
category members at various levels of abstraction. Their
subjects rated the similarity
of
pairs of objects, with the
object pairs differing in taxonomic proximity. The pairs
were from the same subordinate concept, from the same
basic concept (but different subordinates), from the same
superordinate (but different basic concepts), or from com-
pletely unrelated superordinates. Similarity increased from
the general to more specific categories, but the most dra-
matic increase in similarity occurred between the superor-
dinate and the basic levels. That is, two objects in the same
basic category were much more similar than two objects
merely in the same superordinate category. Being in the
same subordinate category added little similarity over be-
ing in the same basic category, leading Mervis and Crisafi
to the conclusion that basic concepts are the most differen-
tiated. That is, only basic categories have both high within-
category similarity and low between-category similarity.
Superordinates lack the first and subordinates lack the
second.
In Experiment 2A, we replicated Mervis and Crisafi's
(1982) method, with one exception. Mervis and Crisafi
used pictures
of
objects, but because one cannot take a
snapshot
of
an event, we composed stories about our
events based on subjects' action lists from Experiment 1.
These stories were the stimuli for the similarity ratings.
If
there is a basic level of events, then the differentiation
scores derived from this procedure should converge with
the data
of
Experiment 1 to reveal the same level as be-
ing basic.
judgments may take into account the qualitative differ-
ences in information that are hidden in feature lists.
412 MORRIS AND MURPHY
similarity and between-eategory similarity. Almost as sur-
prising is the finding that basic and subordinate categories
have about the same differentiation scores.
Experiment 2B
Judging Event Names
One objection to our methodology might be that we are
measuring the superficial similarity of our stories, not the
underlying similarity
of
events. That is, perhaps some-
thing in the way the stories were written, rather than the
true similarities of the events, caused the pattern of results.
To answer this criticism, we replicated Experiment 2A
without using the stories, asking subjects to judge the
categories only on the basis of their names.
Method
The subjects were 12 Brown University undergraduates, who were
paid for their participation. The only difference between Experi-
ments 2A and 2B was that in Experiment 2B, the subjects were given
pairs of subordinate event names, such as quick breakfast and
family dinner, and were asked to rate their similarity. Because they
did not have to read entire stories, each subject rated all 90 pairs
of
names. This procedure created aproblem for measuring the
similarity within the lowest level categories, because there were
no names at a still lower level that could be compared (i.e., no mea-
sure
of
within-category similarity). In order to replicate the previ-
ous experiment as closely as possible, we simply presented the
subordinate name twice in this condition, asking the subjects on
such trials to judge how similar all the events having that name were
(e.g.,
how similar are all English classes?). Although this was a
novel task, the subjects found it perfectly interpretable, and the
results were quite regular. (In any case, the most important ques-
tion being addressed was the relative differentiation of the basic
and superordinate categories because of the unexpected result of
the previous experiment. That comparison was not affected by the
subordinate testing procedure.)
Results and Discussion
Table 3 shows the mean similarity ratings for event
names. The different levels were again reliably different
[F(3,33) =258.25, p<.001]. The striking similarity
of these results with the results from Experiment 2A
demonstrates that the effect stems not from the surface
properties of the stories but from the actual properties
of
the event concepts. The middle-level concepts did not
maximize differentiation; instead, the superordinate level
again was the best in this regard, with a score
of
3.18,
as compared to 1.43 for both the basic and the subordinate
categories. Whether subjects read stories or simply used
memory to judge similarity, the results were rather differ-
ent from comparable measures
of
object concepts. In
short, the results
of
Mervis and Crisafi (1982) were not
replicated with events.
The similarity measures reveal that, despite the dearth
of
features listed for them in Experiment 1, superordinate
concepts carry salient information. Note that this experi-
ment did not disconfirm the predicted informativeness of
middle-level categories. Rather, the results merely show
that superordinates are considerably more informative
than the feature listings indicate. We postpone interpre-
tation
of
this finding until the General Discussion.
EXPERIMENT
3
Objects are often first seen or recognized as members
of
their basic-level category (Jolicoeur, Gluck, &Koss-
lyn, 1984; Murphy &Brownell, 1985; Smith, Balzano,
&Walker, 1978).
If
events have a basic level, subjects
should be able to verify category membership at that level
more rapidly than at the superordinate or subordinate
levels. This should be true even though superordinates
have fewer features than basic concepts (which would
seem to make verification
of
them easier). In Experi-
ment 3, we tested this prediction for event concepts.
Unfortunately, it is impossible to use the same metho-
dology for events as is used for object concepts. As in
Experiment 2, we could not use photographs because, un-
like objects, events cannot be depicted unambiguously in
static pictures. Moreover, we could not use videotaped
events or complete stories, because depicting a subor-
dinate could require as many as 16 actions (see Table 2).
In such stimuli, it would be difficult to prevent informa-
tion about subordinate category membership from appear-
ing later than information about more general categories.
Therefore, the stimuli used were phrases describing parts
of
events rather than entire instances of events. The sub-
jects judged whether an action (taken from the Experi-
ment 1 lists) belonged to an event category-subordinate,
basic, or superordinate. Although this paradigm deviates
somewhat from the familiar object-categorization task, it
should still reveal at which level events are recognized.
In normal experience, events can be categorized as they
unfold, on the basis of their parts rather than on the whole
instance. Thus, we expected categorization reaction time
(RT) to be fastest at the basic level.
Method
Materials.
For each subordinate event, two actions were selected
from the Experiment I action lists. The most distinctive actions were
chosen for each event; we tried to find actions that are definitely
part
of
that subordinate event but definitely not part of the three
contrasting subordinates.
For
two
of
the nine taxonomies (hygiene
and sports), nearly all of the actions of one subordinate also fit con-
trasting subordinates, so only the remaining seven taxonomies were
used. However, it should be noted that some of the actions from
one subordinate seemed plausible actions for the other subordinate
and basic categories, even though the subjects in Experiment I had
not listed them for those categories. (This problem arises because
even though some actions are only typical or expected actions for
one category, they are possible actions for similar categories. Thus,
it becomes difficult to find indisputably false items.) This problem
involves the construction of false trials, whereas the true trials were
of
primary interest. Each action was paired with its three
"true"
categories (one at each level) and with three categories that were
most related while still being
"false,"
resulting in 336 total items.
Thus, for the action
"scream
during the scary
parts,"
the true
categories were horror movie (subordinate), movie (basic), and en-
tertainment (superordinate), and the false categories were western
movie (subordinate), theatre (basic), and transportation (superor-
dinate). Because any of the superordinate categories (except the cor-
rect one) could be a false item, superordinates were randomly as-
signed to false actions such that each one occurred equally often.
The 336 items were randomly split into two lists of 168 items
such that every category appeared in every condition once (thus,
all conditions appeared equally often within each list). Half of the
subjects received one list and half received the other. The order
of trials was randomly determined separately for each subject.
Subjects. The subjects were 16 Brown University undergraduates.
Procedure. The subjects were told that an event name would ap-
pear on the computer screen and that after they had read and un-
derstood it, they should press either one of two response buttons.
The event name would disappear, and (200 msec later) an action
would appear. The subjects were to decide whether that action was
part of the named event. If so, they were to press the button marked
"TRUE";
if not, they were to press the button marked
"FALSE."
The subjects used the dominant hand for the TRUE response and
the nondominant hand for the FALSE response. After a 2-sec de-
lay, the next trial began. Times under 200 msec or over 5 sec were
deleted as outliers or buttonpress errors.
Results
and
Discussion
The true trials (shown in Table 4) led to the predicted
results, but false trials were distorted by the problem
(described above)
of
overlapping actions. For true trials,
the subjects were fastest at the basic level, with the subor-
dinate level close behind and superordinates quite a bit
slower. The three levels were reliably different overall
[F(2,30) =9.61, P<.001 in the subject analysis,
F(2,54) =4.29, P<.025 in the item analysis]. The
subordinate level was not significantly different from the
basic level (both Fs <1), but the superordinate level was
slower
[F(1,30)
=18.99,p <.001, andF(1,54) =8.52,
p<.01]. Error rates increased slightly with category
level [F(2,30) =7.40, P<.005, and F(2,54) =4.49,
P<.025], although this may reflect the greater inclu-
siveness
of
higher level categories rather than a mistaken
decision (i.e., they may not be errors). Overall, this is
apattern quite similar to results for object categories, in
which superordinates are quite slow and subordinates are
nearly as fast as basic categories (e.g.,
see
Jolicoeur et al.,
1984; Murphy & Brownell, 1985; Murphy & Smith,
1982).
For false trials, the most obvious result was that the
subjects did not often agree that the item was false: only
43% and 51 %
of
the subordinate and basic responses,
respectively, were judged false. As noted above, this
reflects the difficulty of finding truly distinctive features
in the feature lists. Since superordinate items were in fact
false 90%
of
the time, there were highly reliable error
effects (though whether these judgments are really errors
is very doubtful) [F(2,30) =199.81, P<.001, and
F(2,52) =32.18, P<.001]. The false RTs show the cor-
responding pattern: long RTs for subordinate and basic
Table 4
Latency and Accuracy of Categorization in Experiment 3
Level of Abstraction
Superordinate Basic Subordinate
Mean Reaction Time 1,481 1,281 1,305
Percentage of
"Yes"
Judgments 79.7 84.8 89.5
Note-Mean reaction times are expressed in milliseconds.
EVENT CONCEPTS 413
trials and short RTs for superordinates [1,610, 1,602, and
1,366 msec, respectively; F(2,30) =12.71, p<.001,
and F(2,52) =7.46, P<.005]. (Because of missing
data, one item had to be eliminated from the false RT anal-
yses.) These slow RTs also undoubtedly reflect the fact
that these actions were not clearly false for subordinate
and basic concepts. In short, because
of
the problem in
finding convincingly false items, the results from the false
trials are probably not fully interpretable. Because so
many of the false items were difficult (not clearly true or
false), the subjects were probably more conservative in
responding true (McCloskey &Glucksberg, 1979). There
is no reason to believe that this could account for the ob-
served basic-level advantage. In fact, since the false trials
were easiest at the superordinate level, this should have
aided true decisions at that level-because its true and false
items would have been the most discriminable. Nonethe-
less, the superordinate true RTs were slowest of all.
The RT advantage of basics over subordinates in true
trials was not statistically significant as it was in Rosch
et al. 's (1976) results (Experiment 7). Yet, since other
experiments with object concepts have shown that this ef-
fect occurs only under very particular conditions (Mur-
phy &Brownell, 1985), this failure to replicate Rosch
et
al.'s
result is unsurprising. The results of the present
experiment, then, support the existence of a basic level
of event concepts insofar as they are comparable with
prior studies. The true trials followed the results of Mur-
phy and Smith (1982), Jolicoeur et al. (1984), and Mur-
phy and Brownell (1985). Because
of
the difficulty in
creating false items, it is somewhat unfair to compare our
false results to those of object-eategorization experiments.
However, this fact raises the rather puzzling question of
why it was so difficult to devise false items for our stimuli
(recall that this problem arose even after two hierarchies
were eliminated because of the overlap problem). One
possible reason is that in object-categorization experi-
ments, a picture of the entire object is presented, whereas
only one part of each event was presented in this experi-
ment. Perhaps if object-identification tasks used single
parts, such as
wings.
seat, or handle, similar results would
obtain. We suggest other possibilities in the General
Discussion.
EXPERIMENT
4
Expanding on Brown's (1958) observations, Rosch
et al. (1976) found that basic-level names are almost
universally preferred in object naming. They explained
this striking effect by positing that basic concepts center
on naturally occurring correlations of features in the per-
ceived world, and argued that communication is easiest
when terms are chosen that exploit these clusters of in-
formation.
If
the level
of
event terms identified by Ex-
periment 1 is truly a basic level, these terms should be
preferred as names for events.
414 MORRIS AND MURPHY
Experiment 4A
Naming Events
To study the terms most useful in communication,
researchers have relied primarily on the free-naming task.
In this procedure, subjects name pictured objects or
scenes, and the terms they use are compared to names
in the taxonomy. Although names at many levels of de-
tail could be used, subjects consistently choose basic-level
terms (e.g., see Berlin, Breedlove, &Raven, 1973). The
basic-name preference is also found in other, more natural
linguistic settings (Brown, 1958; Downing, 1980; Lu-
cariello &Nelson, 1986; Wisniewski &Murphy, 1989).
Once again, the picture-naming task used in earlier in-
vestigations was adapted in the present experiment to a
story-naming task. The terms subjects chose to name the
events in the stories was compared to the taxonomic event
terms.
Basic-level names have been described as the terms of
reference appropriate in neutral or average contexts
(Cruse, 1977). However, names used in definite refer-
ence can change considerably with context: people often
choose names at the level of detail sufficient to distinguish
the referent from other things
around-that
is, from its
contrast set (Olson, 1970). Somewhat surprisingly, Rosch
et al. (1976, Experiment 10) found that basic category
names are preferred across
contexts-not
just in neutral
contexts. That is, even when all the objects in a stimulus
set could be distinguished by superordinate terms, sub-
jects used basic-level terms. And when subordinate names
were necessary to distinguish each object, subjects still
predominantly used basic category terms. To investigate
whether shifting the context affects preference for basic-
level names
of
events, we also compared three contrast
sets.
Method
Stimuli. The 72 stories from Experiment 2 were used as stimuli.
These stories were derived from the subordinate categories, which
ensured that terms at all three levels could appropriately name the
event described. Recall that the subordinate event term itself oc-
curred in the first sentences of the story, since the subjects in Ex-
periment Ihad listed actions in that context. Although this pro-
cedure may bias subjects toward using the subordinate name, it
ensured that the event could be identified at that level.
The contrast set conditions were created by varying the similar-
ity of the stories in the set. The superordinate contrast set contained
no stories from the same superordinate categories, so the stories
could be distinguished with superordinate terms. The basic con-
trast set contained stories from the same superordinate but not from
the same basic categories, so that basic terms were sufficient to
distinguish them. The subordinate contrast set contained stories from
the same basic categories, so that subordinate terms were required
to distinguish them. This manipulation was designed to discover
whether subjects would use the most general name that would dis-
tinguish the stories (Olson, 1970) or instead use the basic level
regardless of context, as Rosch et aI. (1976) found. For the superor-
dinate contrast condition, two sets
of
nine stories (one from each
superordinate category) were created. For the basic contrast con-
dition, one set of 18 stories (one from each basic category) was
created. For the subordinate contrast condition, two sets were
created-one
with all 4 subordinate stories from each of the first
4taxonomies (16 stories), the other with all 4 subordinates from
the remaining 5 taxonomies (20 stories). For all five sets, there were
two versions: one in a random order and the other in the reverse
order.
Subjects. The subjects were 40 Brown University undergradu-
ates, who were paid to participate. Each set of stories was named
by 8 subjects (4 received the forward version, and 4 received the
reverse version), so that an equal number of subjects saw each story
in each contrast set condition.
Procedure. The subjects were first instructed to read through
the sets of stories quickly. This procedure ensured that they were
aware of the contrast set before naming. They were then instructed
to read through the set again, this time naming the event in each
story with a short phrase. To discourage idiosyncratic, artistic ti-
tles, the instructions borrowed Tversky and Hemenway's (1983)
injunction to give
"the
most simple, obvious, direct sort of name
that ordinary people would give for each
event."
Item tally. Ajudging procedure was used to classify responses
as superordinate-, basic-, or subordinate-level names. Judges (an
author and an independent judge) were given scoring sheets that
listed the source subordinate term and its basic and superordinate
terms for each story. The judges were instructed to
"match
each
name on the answer sheet to one of the three on the scoring sheet
...
choosing the name on the score sheet that resembles the sub-
ject's
answer most closely in level of detail or generality." The
two judges agreed on
90%
of the names, and a third independent
judge resolved the conflicts.
Results
This experiment was designed to see whether subjects
would use basic-level names to label events, even when
more specific or general names would be appropriate. Ta-
ble 5 shows the number
of
names from each taxonomic
level used by the subjects across contrast set conditions.
Clearly, the basic-level names were favored, and superor-
dinate names were hardly ever used. The differences be-
tween the three levels were reliable
[X
2(2) =230.22,
p<.00 1]. A second question is whether contrast set af-
fects the level of detail at which events are named. A test
of
independence found that the number
of
names at each
level did depend on the contrast set fx2(4) =31.83,
p<.001]. As Table 5 clearly shows, superordinate
names were seldom used in any
condition-even
when
they were sufficient to distinguish the stories. But subor-
dinate names were used far more often when they were
necessary to distinguish the stories than in other condi-
tions. Thus, the basic-level advantage was not completely
independent of contrast set, contrary to what Rosch et al.
(1976) found for object names. Nonetheless, basic
category names were used most often in all conditions,
showing that the subjects were not simply choosing the
names that would give an unambiguous reference (see
Clark &Murphy, 1982).
Table 5
Number of Names Used at Each Category Level in Experiment 4
Level of Abstraction
Contrast Set Superordinate Basic Subordinate
Superordinate 12 102 30
Basic 3108 33
Subordinate 18 140 130
Total 33 350 193
EVENT CONCEPTS 415
Table
6
Mean Appropriateness Ratings from Experiment 4
Method
Afterthe
subjects
finished
Experiment
4A, their
answer
sheets
were
collected,
and
they
were
instructed
to
read
the
stories
again
quickly.
The
subjects
were
then
given
rating
sheets
with
the
three
possible
names
foreach
story
in their set.
Thus,
corresponding
to
the story
about
going
to a horror
movie,
there
appeared:
Entertainment__
Movies__
Horror
Movies
__
The
instructions
asked
for
ratings
of the "appropriateness of
each
name
for the event" on a
scale
of 1-10,
where
I
means
"not at
all appropriate" and 10
means
"most appropriate."
Results and Discussion
This experiment was designed to discover whether sub-
jects consider subordinate-level terms appropriate names
for the events in the stories, even though they chose basic-
level terms more often in the free-naming task. In fact,
the subjects rated subordinate terms as more appropriate
names for the stories than basic-level terms. Table 6
shows mean appropriateness ratings for names at each
level in each contrast set condition. Superordinate terms
were rated much less appropriate than the other two levels,
and the effect of category level was highly reliable
[F(2,16) =157.65,p <.001]. The subjects undoubtedly
Experiment 48
Rating Event Names
Before drawing any conclusions about the role of basic-
level event concepts in communication, some alternative
explanations must be ruled out. It may be that subjects
avoid subordinate-level terms because they are unfamiliar.
Another possible interpretation is that subjects avoid
specific names because details
of
the stories deviate from
their experiences.
If
asubject had attended horror mo-
vies only at drive-in theaters, he or she might consider
"horror
movies"
an inappropriate name for the story
about ahorror movie at a walk-in theater. For that sub-
ject, the most specific name fitting the story would be the
basic-level term
"movies."
To test both of these critical
assumptions-that
the subordinate terms are familiar to
the subjects and appropriate to the
stories-we
conducted
a final experiment in which subjects rated the appropri-
ateness
of
all the taxonomic terms for the stories. We ex-
pected that the subordinate names would be rated at least
as appropriate as the basic-level names, thus ruling out
these alternative explanations for why subjects produced
basic names in Experiment 4A. In fact, because the subor-
dinate names describe the stories most precisely, they may
be judged the most appropriate.
felt that subordinate terms were the most appropriate
names for the stories, but these were not the names spon-
taneously chosen in most conditions. Although there was
a main effect of contrast set[F(2,16) =
21.76,p
<.001],
and the interaction between taxonomic level and contrast
set was significant [F(4,32) =3.14, P<.05], these ef-
fects do not mitigate the conclusions, since the subor-
dinates were rated more appropriate than the basic terms
in every contrast condition. These results justify the as-
sumption that the subjects used basic-level names (in Ex-
periment 4A) even when subordinate names would have
been fully appropriate.
Experiment 4 showed that basic-level event names are
greatly preferred in naming. Superordinates were virtu-
ally never used, and subordinates were used only in the
subordinate contrast condition, where they were neces-
sary to distinguish the stories. But even in this condition,
the subjects used basic names about as often as subordinate
names. Nonetheless, this shift with contrast set was one
difference between event naming and object naming
reported by Rosch et al. (1976). One possible explana-
tion of this difference is that although objects tend to have
familiar nouns that denote their basic-level name, many
events require multiple-word descriptions (e.g., going to
the theater, seeing a play). Often these descriptions are
not standardized, in that any of a number of phrases could
be used, with none having clear priority. Perhaps the con-
trast set has a stronger influence on event naming because
of the lack of conventional, standard names. Overall,
event terms at the basic level are clearly the preferred
name in neutral settings, and their superiority is only
somewhat abated in other contexts.
The rules of cooperative communication may explain
why the structural characteristics of the basic level lead
subjects to prefer basic category names (Clark &Mur-
phy, 1982; Cruse, 1977). One commonly accepted con-
versational rule is for speakers to be informative but not
to give unnecessary information (Grice, 1975). Labeling
an event with a superordinate name (e.g., entertainment)
provides insufficient information about the event. As Ta-
ble 2 shows, most of the concrete information about an
event is given by the basic level. Superordinate names
may be used when referring to a number of events at once
(i.e., no single event may be called "entertainment," but
acollection of events might be; Wisniewski &Murphy,
1989). Subordinate names may only be used when the
specific information associated with that level is particu-
larly relevant (Clark &Murphy, 1982). Even when the
subordinate was necessary to distinguish an event from
its contrast set, it was not the name that most subjects
chose, probably because the marginal gain in informa-
tion from moving to the subordinate was so small.
Unlike the naming results, appropriateness ratings were
a direct function of the specificity of the name, with subor-
dinates receiving the highest rating and superordinates
receiving the lowest ratings. This pattern is very similar
to that found in object concepts by Smith and Osherson
(1984; Smith, Osherson, Rips, &Keane, 1988). They
9.12
9.47
9.64
9.41
Subordinate
7.66
8.53
788
8.02
3.93
4.52
4.96
4.47
Level
of
Abstraction
Superordinate
Basic
M
Contrast
Set
Superordinate
Basic
Subordinate
416 MORRIS AND MURPHY
found that an object was judged most typical of the most
specific category that it fit (e.g., a red apple would be
judged more typical of red apple than
of
apple). Their
theory explains this result as a function of the number of
shared features in the object and concept representation.
That is, a red apple shares more features with the con-
cept red apple than it does with apple or fruit. Asimilar
account would explain the present results as well: The
more specific an event concept, the more closely it will
match the story, and the higher the appropriateness rat-
ing will be. Most importantly, the processing advantages
of
basic-level concepts cannot be explained away as
reflecting only appropriateness or typicality.
GENERAL DISCUSSION
In this study, we repeated procedures used to identify
the basic level of object concepts to investigate whether
concepts for routine social events have basic-level struc-
ture. The general pattern of results from our four experi-
ments seems to replicate the pattern of informativeness
and processing advantages marking the basic level.
Against the general trend of convergence, however, were
two notable divergences: the finding that superordinate
event concepts were very differentiated in similarity rat-
ings (Experiment 2) and the shift in names chosen toward
the subordinate level in contrast sets consisting of highly
similar events (Experiment 4). One purpose in studying
concepts for entities other than material objects is to iso-
late the properties
of
basic-level structure from the proper-
ties of the object domain. We interpret the similarities in
results for objects and events as evidence for a common
organizational
structure-the
basic
level-across
domains,
and we interpret the differences as indications
of
how
unique properties of each domain interact with the com-
mon structure.
Extending Rosch's method and theory from the object
domain to entities in other ontological domains allows us
to extend some of her general conclusions. There are two
that we would like to highlight before discussing the de-
tails
of
event concepts. First, it seems that the structure
of
the world
of
events (as perceived by human agents)
affords optimal categorization at an intermediate level of
abstraction. This is because attributes are not distributed
orthogonally or randomly across events; many attributes
covary, and categories at an intermediate level can center
on clusters of co-occurring attributes. Second, social event
concepts serve cognitive purposes, as do object concepts,
such that greater differentiation leads to an advantage in
processing. These general conclusions are supported by
the pattern of specific results described in the next section.
Event Representations
To review the interaction of basic-level organization and
the unique properties of events, we propose an account
of
how event information, from abstract to concrete, is
represented in concepts. To begin our discussion, let us
first describe superordinate event concepts (e.g., enter-
tainment). We suggest that, like object superordinates,
superordinates carry information about the functions, pur-
poses, and goals
of
activities (Tversky &Hemenway,
1984). Categorizing at this level highlights abstract com-
monalities between activities that have a similar role in
our culture. The coherence
of
these concepts comes not
from below, the sharing
of
perceivable features, but from
above, their common role in
"theories"
of
human be-
havior (Murphy &Medin, 1985). Hence, few features
are given for superordinates in lists
of
attributes (Rifkin,
1985) or lists of actions (Experiment 1). Yet, because of
their similar functional roles, these categories are coher-
ent: events within the same superordinate are far more
similar to each other thanto events from different superor-
dinates (Experiment 2), even though they do not share
the same parts.
For
example, participants in all forms of
entertainment share common goals and feelings, even if
they differ in their actual actions. Since superordinate
terms are not directly associated with specific actions, they
are slower in the categorization task (Experiment 3). Fi-
nally, because superordinates do not provide concrete de-
tails about individual events, they are not used as names;
more specific concepts are so used (Experiment 4).
Basic-level event concepts (e.g., going to the movies)
carry the abstract functional information of superordinates
but also a wealth of concrete, perceptual information. The
critical features that make the basic level so primary may
be parts (Tversky &Hemenway, 1984), which are func-
tional and perceptual divisions
of
the event. When sub-
jects list the parts
of
events (actions), the informativeness
of basic-level concepts is dramatically revealed (Experi-
ment 1). Other kinds
of
features such as objects, loca-
tions, times, and outcomes, also cluster at the basic level
(Rifkin, 1985). This information makes basic concepts
more coherent than superordinates (Experiment 2). Be-
cause basic-level concepts are strongly associated with
specific actions, they are used more quickly in categori-
zation than are concepts at other levels (Experiment 3).
And because basic-level concepts are both distinctive and
informative, they are generally chosen to name events
(Experiment 4). The informativeness and processing ad-
vantages shown by concepts at the basic level suggest that
memory for events may be primarily encoded at this level.
In other words, these concepts are more available than
more abstract or more concrete alternatives.
Individually, subordinate event concepts (e.g., going
to a horror movie) are much like basic concepts with a
bit more detail. Collectively, they differ in that they are
more similar to each other, giving them little distinctive-
ness. That is, going to a horror movie is very much like
going to a western. Furthermore, it seems likely that the
distinct features
of
subordinates tend not to be parts of
the event, but are perceptual and contextual details. For
example, Experiment 1 found that, on average, subor-
dinates added less than one new action to the actions that
make up basic categories. Yet subordinates were con-
sidered by the subjects in Experiment 2 to have consider-
ably more within-eategory similarity than basic categories,
so there must be other aspects (besides parts) of subor-
dinates that make them similar. These aspects may include
the manner
of
carrying out an action (as for quick break-
fast), the objects involved (assault with a deadly weapon),
or the location (Amtrak to New York), rather than entire
parts. In categorization tasks, subordinates can be used
no more quickly than basic concepts (Experiment 3),
presumably because of their lack of distinctiveness. Subor-
dinates are not preferred in naming generally, but in con-
texts in which their details are relevant for distinguishing
events, subjects do use their names (Experiment 4).
Differences Between Object and Event Concepts
The description
of
event concepts that we have just pro-
posed is in many respects quite similar to theories of levels
of
object concepts. Although there are some important
caveats in this correspondence, it seems that some aspects
of
conceptual structure are universal enough to apply both
to events and to
objects-two
very different ontological
types. Nonetheless, there are some important differences
between the two, and the rest
of
the discussion focuses
on these differences.
One difference between object and event categories is
the apparently different representations of superordinates.
Experiment 2 found that event superordinates are highly
informative, although in Experiment I few features were
listed for them. The key to this paradox may lie in qualita-
tive rather than quantitative differences between features.
Object superordinates carry functional, not perceptual, in-
formation (Tversky &Hemenway, 1984). And Rifkin
(1985, p. 542) noted that event superordinates contain a
higher proportion of "abstract person-related features,"
such as
"competition"
and
"fun"
for sports. These ab-
stract features seem more central to event concepts than
to object concepts, since events are structured by goals
rather than by a perceptual shape.
For
example, what
makes an activity entertainment is not some specified set
of physical actions, but rather the fact that the participants
are doing it in order to have fun. Similarly, Read, Jones,
and Miller (in press) have argued that goals are central
to personality concepts, and Hampson, John, and Gold-
berg (1986) have found that personality superordinates
are relatively more informative than object superordinates.
We noted above that events often do not have ready-
made names for them, as objects do. When someone asks
you what you are doing, there is often no single name
that is the conventional label for that activity. One might
easily respond with a number of names that focus on
different aspects
of
the activity, at different levels of ab-
straction and including more or fewer actions. This might
explain why, although there is a basic-level naming ad-
vantage (Experiment 2), it is not as stable across contexts
as in object hierarchies.
This difference between the codabilities of objects and
events could be an artifact of the English language, but
there is reason to believe that event categorization is in-
trinsically less stable than object categorization. Gentner
(1982) analyzed the componentiality and boundedness of
EVENT CONCEPTS 417
actions versus objects in order to explain differences be-
tween nouns and verbs. Languages differ greatly in what
aspects of an action are included in a single verb, whereas
they largely agree on what counts as an individual noun
referent. This occurs, Gentner argued, because actions
have a number of components that are in fact separated
in different actions. An action might include an actor, an
object, a recipient, and an activity, done in a certain man-
ner at a certain time in a certain location or context.
However, at other times, the same activity might be done
with a different actor, in a different manner, and so on.
Different verbs (and different languages) encode differ-
ent components of these actions, which suggests that there
is no single, preferred way to conceptualize every action.
The parts of objects, on the other hand, are often not phys-
ically separable from each other and do not occur alone.
Equally important, the boundaries
of
objects are much
more clearly defined than those of actions. It is often very
difficult to delineate precisely when an event has begun
and ended. Does a party begin with the preparations, the
arrival of the first guest, or when all the guests have ar-
rived? Parties, like many events, often have no well-
marked beginning and ending. In contrast, there is little
difficulty in discerning where a chair ends and the floor
begins; the boundaries of objects are usually very clear.
Another reason that object categories have greater sta-
bility is that for some kinds
of
objects, physical laws im-
pose a categorical structure. For example, natural kinds,
such as chemical elements or biological species, are
shaped by natural processes and seem to exist as categories
prior to human activities (Putnam, 1970). In contrast,
categories of social events are necessarily shaped by hu-
man goals. Event categories tend to be more heterogen-
eous and overlapping because human goals are plastic and
multiplex. Vallacher and Wegner (1985,1987) have ana-
lyzed how actions can be identified on many different
teleological levels. For example, a single action can be
categorized as pressing buttons, making a phone call, or
initiating afriendship. Thus, the conceptualization of ac-
tions is more variable. This variability was seen in Ex-
periment 3, where it was difficult to find actions that were
true of one category but clearly false of related categories.
However, this is not to say that there is no variability in
object categorization, as Barsalou's (1983, 1987) work
has demonstrated.
We have emphasized that there may be many different
ways
of
classifying events. Cognitive psychologists often
assume a taxonomic hierarchy for concepts because of its
property of inferential inheritance (Collins &Quillian,
1
969)-that
is, properties of high-level categories are also
properties of their subordinates. However, other students
of
events have focused on entirely different classification
systems, such as partonomies (Abbott et al., 1985; Reiser,
Black, &Abelson, 1985). For example, they might com-
pare going to a restaurant with ordering food and telling
the waiter your selection. And Vallacher and Wegner
(1985, 1987) have discussed event concepts as organized
by networks of goals. Script organizations, based on the
418 MORRIS AND MURPHY
work of Schank and Abelson (1977), do not support in-
ferential inheritance (e.g., if going to a restaurant makes
you less hungry, it does not follow that the act of order-
ing makes you less hungry). Nonetheless, these organi-
zations are crucial to the planning of activities and the
understanding of the course of events in some activity.
So an ordinary understanding of events probably includes
a partonomic structure (as in script theory), ahierarchi-
cal structure (as studied in this article), and various goal-
oriented relations. Now that the basic level in taxonomies
has been identified, the question of whether it plays a spe-
cial role in these other organizations can be addressed.
REFERENCES
ABBOTT,
V.,
BLACK,
J. B., &
SMITH,
E. E. (1985). The representation
of scripts in memory. Journal
of
Memory &Language, 24, 179-199.
ADELSON,
B. (1983). Constructs and phenomena common to
semantically-rich domains (Cognitive Science Tech. Rep. No. 14).
New Haven, CT: Yale University.
BARSALOU,
L. W. (1983). Ad hoc categories. Memory &Cognition,
11,211-227.
BARSALOU,
L. W. (1987). The instability of graded structure: Implica-
tions for the nature of concepts. In U. Neisser (Ed.), Concepts and
conceptual development: Ecologicaland intellectualfactors in categori-
zation (pp. 101-140). Cambridge: Cambridge University Press.
BARSALOU,
L. W., &
SEWELL,
D. R. (1985). Contrasting the represen-
tations of scripts and categories. Journal
of
Memory &Language,
24, 646-665.
BERLIN,
B.,
BREEDLOVE,
D. E., &
RAVEN,
P. H. (1973). Generalprin-
ciples of classification and nomenclature in folk biology. American
Anthropologist, 75, 214-242.
BoWER,
G. H.,
BLACK,
J. B., &
TURNER,
T. J.
(1979).
Scripts
in
memory
for text. Cognitive Psychology, 11, 177-220.
BROWN,
R. (1958). How shall a thing be called?Psychological Review,
65, 14-21.
CANTOR,
N., &
MISCHEL,
W. (1979). Prototypes in person perception.
In L. Berkowitz (Ed.), Advances in experimental social psychology
(Vol. 12, pp. 3-52). New York: Academic Press.
CANTOR,
N.,
MISCHEL,
W.,
8<
SCHWARTZ,
J. C. (1982). A prototype
analysis
of psychological situations. Cognitive Psychology, 14,45-77.
CLARK,
H. H., &
MURPHY,
G. L. (1982). Audiencedesign in meaning
and reference. In J. F. Le Ny & W. Kintsch (Eds.), Language and
comprehension (pp. 287-299). Amsterdam: North-Holland.
COLLINS,
A., &
QUILLIAN,
M. R. (1969). Retrieval time and semantic
memory.Journal
of
Verbal Learning &Verbal Behavior, 8, 240-247.
CRUSE,
D. A. (1977). The pragmatics of lexical specificity. Journal
of
Linguistics, 13, 153-164.
DOWNING,
P. (1980). Factors influencing lexical choice in narrative.
In W. L. Chafe (Ed.),
The
pear stories: Cognitive, cultural, and lin-
guistic aspects
of
narrative production (pp. 89-126). Norwood, NJ:
Ablex.
GENTNER,
D. (1982). Why nouns are learned before verbs: Linguistic
relativity versus natural partitioning. In S. A. Kuczaj II (Ed.), Lan-
guage development: 2. Language, thought and culture (pp.
301-304).
Hillsdale, NJ: Erlbaum.
GRICE,
H. P. (1975). Logic and conversation. In P. Cole & J. L. Mor-
gan (Eds.), Syntax and semantics: 3. Speech acts (pp. 41-58). New
York: Academic Press.
HAMPSON,
S. E.,
JOHN,
O. P., &
GOLDBERG,
L. R. (1986). Category
breadth and hierarchical structure in personality: Studies in asym-
metries in judgments of trait implications. Journal
of
Personality &
Social Psychology, 51, 37-54.
HAMPTON,
J. A. (1987).Inheritance of attributes in naturalconceptcon-
junctions. Memory &Cognition, IS, 55-71.
JOLICOEUR,
P.,
GLUCK,
M., &
KOSSLYN,
S. M. (1984). Pictures and
names: Making the connection. Cognitive Psychology, 16, 243-275.
LUCARIELLO,
J., &
NELSON,
K. (1986). Contexteffects on lexicalspeci-
ficity in maternal and child discourse. Journal
of
Child Language,
13, 507-522.
MCCLOSKEY,
M., &
GLUCKSBERG,
S. (1979). Decision processes in
verifying category membership statements: Implicationsfor models
of semantic memory. Cognitive Psychology, 11, 1-37.
MERVIS,
C. B., 8<
CRISAFI,
M. A. (1982). Order of acquisition of
subordinate-, basic-, and superordinate-level categories. Child De-
velopment, 53, 258-266.
MURPHY,
G. L., &
BROWNELL,
H. H. (1985). Category differentiation
in object recognition: Typicalityconstraints on the basic categoryad-
vantage.Journal
of
Experimental Psychology: Learning, Memory, &
Cognition, 11, 70-84.
MURPHY,
G. L., &
MEDIN,
D. L. (1985). The role of theories in con-
ceptual coherence. Psychological Review, 92, 289-316.
MURPHY,
G. L., &
SMITH,
E. E. (1982). Basic-levelsuperiority in pic-
ture categorization. Journal
of
Verbal Learning &Verbal Behavior,
21, 1-20.
OLSON,
D. R. (1970). Language and thought: Aspects of a cognitive
theory of semantics. Psychological Review, 77, 257-273.
PUTNAM,
H. (1970). Is semantics possible? In H. E. Kiefer & M. K.
Munitz
(Eds.), Language, belief, andmetaphysics(pp.
50-63).
Albany,
NY: State University of New York Press.
READ,
S. J.,
JONES,
D. K., &
MILLER,
L. C. (in press). Traits as goal-
based categories: The importance of goals in the coherence of dis-
positional categories. Journal
of
Personality &Social Psychology.
REISER,
B. D.,
BLACK,
J. B., &
ABELSON,
R. P. (1985). Knowledge
structures in the organization and retrieval of autobiographical
memories. Cognitive Psychology, 17, 89-137.
RIFKIN,
A. (1985). Evidence for a basic level in event taxonomies.
Memory &Cognition, 13, 538-556.
ROSCH,
E. (1978). Principles of categorization. In E. Rosch & B. B.
Lloyd (Eds.), Cognition and categorization (pp. 27-48). Hillsdale,
NJ: Erlbaum.
ROSCH,
E.,
MERVIS
C. B.,
GRAY,
W.,
JOHNSON,
D., &
BOYES-
BRAEM,
P. (1976). Basic objectsin natural categories. Cognitive Psy-
chology, 8, 382-439.
ScHANK,
R. C., &