You are on page 1of 26

JOURNAL OF VERBALLEARNING AND VERBALBEHAVIOR 22, 449--474 (1983)

Mental Models of Word Meaning

TIMOTHY P . M C N A M A R A AND ROBERT J. STERNBERG

Yale University
Four experiments compared five alternative mental models of word meaning: a defining-
attributes model, two characteristic-attribute models, and two mixed models. The two char-
acteristic-attribute models (and the two mixed models) differed from each other in terms of
whether weights of attributes were summed or averaged. In Experiment 1, 11 subjects pro-
vided lists of attributes for 24 different nouns. Subjects also rated the necessity, sufficiency,
and importance of each attribute to a given word's definition. Results showed that both de-
fining (necessary and sufficient) and characteristic attributes could be identified from sub-
jects' ratings for most nouns, but that the role of defming attributes was especially pro-
nounced for proper names. In Experiment 2, the same subjects later rated the extent to
which given lists of attributes successfully described referents of each given word. Mathe-
matical modeling of results supported a mixed model of reference based on defming attri-
butes plus the weighted sum of all attributes (defining and characteristic). Experiment 3 rep-
licated the results of Experiment 1 with 25 new subjects and slightly modified procedures. In
Experiment 4, 32 subjects were presented with two alternative attribute lists for each given
word, and were to select the better referent of the word as quickly as possible. The preferred
mixed model from Experiment 2 was shown best to account for both the mental representa-
tion and real-time processing of word meanings.

It has been commonly held by psycholo- how it is that words are "hooked up" with
gists and philosophers that the meaning of a objects in the world. (We in no way claim
word is given by a set of attributes or se- that the three classes of models are exhaus-
mantic markers and that referring consists tive with respect to all possible theories of
of matching attributes of an object with the meaning and reference.) We then describe
attributes or markers included in the word' s four experiments that attempt to distinguish
definition. However, theories of meaning among these three classes of models.
and reference differ with respect to what
constitutes a successful match and how at- MODELS OF M E A N I N G AND REFERENCE
tributes or markers are weighted according Traditional Models
to their importance in the matching pro-
cess. In the introduction to this article, we According to traditional models, the
first outline three classes of models of meaning of a word is decomposed into a
meaning and reference that have been pro- set of attributes (Frege, 1952; Russell,
posed by either psychologists or philoso- 1956) or a set of semantic markers (Glass
phers, or both, to explain what is present to & Holyoak, 1974/1975; Katz, 1972) that
the mind when a word is understood, and are individually necessary and jointly suf-
ficient for determining the word's mean-
This research was supported by Contract ing. 1 That is, the meaning of a word can
N0001478C0025 from the Office of Naval Research
to Robert J. Sternberg, and was conducted while 1 We will use "attribute" in a fairy loose way, en-
Timothy P. McNamara was supported by a Yale compassing with "attribute" what others have called
University Graduate Fellowship. Portions of this ar- "features" (Smith & Medin, 1981), "semantic mark-
ticle were presented at the Twenty-second Annual ers" (Katz, 1972), and "attribute-value pairs"
Meeting of the Psychonomic Society, Philadelphia, (McCloskey & Glucksberg, 1979). For us, an attri-
Pa., 1981. Requests for reprints should be sent to bute is a proposition about the state of the world; it
Robert J. Sternberg, Department of Psychology, is a piece of world knowledge. We make no claims
Yale University, Box 11A Yale Station, New Haven, about how semantic attributes are organized in mem-
Conn. 06520. ory (e.g., networks vs lists).

449
0022-5371/83 $3.00
Copyright© 1983by AcademicPress, Inc.
All rightsof reproductionin any formreserved.
450 MC NAMARA AND STERNBERG

be given by some number of attributes tribute weights (Smith & Medin, 1981) or
common to all exemplars of the word, by averaging attribute weights. To our
which, when jointly possessed by some knowledge, the difference between sum-
object, guarantee that the object is an ming and averaging weights has not been
exemplar of the word. These necessary- addressed in any general theory of word
and-sufficient attributes are often called meaning and reference, although it has
"defining attributes." For example, the been considered in other contexts, such
meaning of " g o l d " might be represented as information integration in impression
as, "is yellow," "is malleable," "is formation (Anderson, 1979).
heavy," and so on. Having all of these at-
tributes is sufficient for something to be Mixed Models
gold, and to be gold, it is necessary that The third class of models of meaning
an object be yellow, malleable, heavy, and reference represents a combination of
and so on. aspects of traditional models and aspects
of characteristic-attribute models. An ex-
Characteristic-Attribute Models ample of a " m i x e d " model is that pro-
According to characteristic-attribute posed by Smith, Shoben, and Rips (1974).
models, the meaning of a word is decom- Smith et al. make the distinction between
posed into a set of attributes that tend to two kinds of attributes in the definition of
be characteristic of exemplars of the a category: defining attributes provide the
word. Some of these attributes can be necessary and sufficient conditions for
necessary, but not all of them have to be category membership, whereas character-
necessary, as in traditional models. A istic attributes determine the prototypical-
well-known example of the usefulness of ity of the members (Rosch, 1973; Rosch
this kind of model stems from Wittgen- & Mervis, 1975). The two kinds of attri-
stein's (1953) analysis of the concept butes are further distinguished by the way
" g a m e . " It is extremely difficult to speak they are used in a two stage categoriza-
of necessary attributes of a game, let tion process. In the first stage, an overall
alone a set of attributes that is both nec- measure of category membership is deter-
essary and sufficient. Nevertheless, we do mined by matching all of the candidate
have a concept " g a m e , " and exemplars of exemplar's features with all of the attri-
this concept seem to share some "family butes in the definition of the category, in-
resemblance" to each other. cluding both defining and characteristic
This class of models can be divided into attributes. If this stage does not result in
at least three submodels according to how an easy decision, a second, more analytic
attributes are used to refer. In the first stage is invoked: Here the candidate ex-
submodel, the referent of a word is deter- emplar and the category are compared on
mined by the number of attributes pos- the basis of defining attributes alone.
sessed by an object that match attributes In the model proposed by Smith et al.
in the word's definition (Hampton, 1979; (1974), all categories have defining sets of
Schwartz, 1977). In the second and third attributes. But this requirement is not a
characteristic-attribute models, it is sup- necessary part of mixed models of mean-
posed that some attributes are more criti- ing and reference. Clark and Clark (1977,
cal than other attributes, and so are p. 467) have briefly discussed a mixed
weighted more heavily (Hampton, 1979; model that does not require that all words
Schwartz, 1977; Smith & Medin, 1981). have defining sets of attributes. According
These two models can be distinguished on to this mixed model, if a word has defin-
the basis of whether the decision process ing sets of attributes, then an object must
accounts for importance by summing at- possess these attributes to be considered
M O D E L S OF M E A N I N G 451

an exemplar of the word. However, if a nouns (see, e.g., Kripke, 1972). Subse-
word does not have defining sets of attri- quent to listing attributes, subjects were
butes, then an object must possess a cer- required to rate each attribute for each
tain number of the characteristic attri- word on a number of scales. These scales
butes in the intension of the word to be measured whether or not the attribute was
considered an exemplar of the word. needed for something to be an exemplar
The experiments reported in this paper of the word (necessity), whether or not
were performed to test versions of the the attribute guaranteed being an exem-
three classes of models of meaning and plar of the word (sufficiency), and finally,
reference. Traditional models and charac- how important the attribute was for defin-
teristic-attribute models have received con- ing the word.
siderable attention in psychological re- The goals of the second experiment
search (see Smith & Medin, 1981, for an were to (a) determine the extent to which
excellent review of this literature). Mixed subjects use defining and characteristic at-
models of semantic representation, on the tributes when referring, (b) test the three
other hand, have been virtually ignored classes of models of meaning and refer-
(but see Smith et al., 1974, for an explicit ence in a rating task, and (c) assess the
formulation of a mixed model). The ex- generalizability of the model fits across
periments that have tested mixed models word domains. Subjects were given a list
have been primarily concerned with pro- of words paired with descriptions of ob-
cessing assumptions made by particular jects, people, or concepts. For example,
mixed models (e.g., the two stage model subjects might see "Tiger" paired with
of Smith et al.) rather than with represen- "Member of the cat family, Has orange
tational assumptions (Glass & Holyoak, stripes, Lives in Asia." The descriptions
1974/1975; Hampton, 1979, 1981; McClos- were constructed from the attributes col-
key & Glucksberg, 1979; Ratcliff & Mc- lected in the first experiment from these
Koon, 1982). One of the goals of the experi- same subjects. A subject's task was to
ments reported here was explicitly to test give a confidence rating that the described
various mixed models of semantic repre- object, person, or concept was an exem-
sentation and processing. In conducting plar of the word paired with the descrip-
such tests, we were also able to test ver- tion.
sions of the traditional- and characteristic- The goals of the third and fourth experi-
attribute models. ments were to (a) verify the results from
The goals of the first experiment were the first and second experiments using
to (a) determine whether or not subjects converging operations, and (b) test a
could find necessary or sufficient attri- model of the real-time processing of at-
butes, (b) assess the degree to which nec- tribute information. Subjects in the third
essary-and-sufficient attributes could be experiment provided ratings similar to
identified from subjects' ratings, and (c) those collected in the first experiment.
collect materials needed for later experi- In the fourth experiment, a timed deci-
ments. We obtained lists of attributes for sion task was used to test alternative
24 nouns, which included names of things models of meaning representation and
occurring naturally in the world (e.g., attribute information processing. Subjects
eagle and diamond) and names of famous were shown a word followed by two lists of
people (e.g., Richard Nixon and Moses). attributes. For example, subjects might see
Proper names were included because they " L a m p " followed by (a) "Electrical, Has a
have been commonly used by philoso- shade," and (b) " H a s a shade, Has a bulb,
phers as a starting point to construct Used for decoration." Their task was to
theories of meaning and reference for all choose the described object, person, or
452 MC NAMARA AND STERNBERG

concept that was the better exemplar of subjects were given a list of the 24 words
the target word. Using subjects' response and asked to list attributes of exemplars
choices and response latencies as the de- of the words. Subjects were allowed one
pendent measures, we hoped to model the minute for each word. Similar procedures
representation and processing of infor- have been used by Rosch and Mervis
mation when making fine judgments of (1975) and by Hampton (1979, 1981). A
meaning. master list of attributes was constructed
from the attributes listed by subjects. At-
EXPERIMENT 1
tributes for this list were chosen from
The first experiment involved collecting subjects' attribute listings in decreasing
semantic attributes and obtaining ratings order of production frequency. Attributes
on those attributes. were chosen for a word until either 10 at-
Method tributes were compiled for the word or
there were no more attributes that at least
Subjects
two subjects listed for that word. Thus,
The subjects were 10 Yale University the numbers of subjects who listed an at-
undergraduates and 1 Yale graduate stu- tribute ranged from 2 to 11. The master
dent. All 11 subjects participated in the list of attributes and the number of sub-
first session of the experiment. One sub- jects who listed each attribute can be
ject was later dropped from the study be- -~,found in the Appendix.
cause he failed to follow directions in one To control for possible individual differ-
of the rating tasks. Subjects were compen- ences in the representation of word mean-
sated for their participation with course ings, a word-attribute list was constructed
credit and monetary payment. for each subject by adding that subject's
unique responses (i.e., listed attributes) to
Materials
the master list. This list could include factu-
Three kinds of words were used as ally incorrect attributes (e.g., that Paul
stimuli: natural kind terms, artifact terms, Newman is dead). In all subsequent ses-
and names of famous people. The natural sions of the experiment, a subject saw
kind and artifact terms were selected from every attribute that he or she listed for each
the Battig and Montague (1969) norms. word and some attributes that he or she did
Eight words were chosen for each word not list (i.e., those attributes that at least
type, with each word a member of a dif- two other people had listed).
ferent category. Word frequency (Ku~era In the second, third, and fourth sessions,
& Francis, 1967) was approximately bal- subjects made three kinds of ratings, each
anced both within and across the natu- on different days. In one session, subjects
ral kind and artifact terms. Seven of were asked to check off those attributes
the proper names were names of famous that exemplars of the words must have to
people who had lived in a number of dif- be exemplars (i.e., necessary attributes). In
ferent time periods, who represented a another session, subjects were asked to
number of occupations, and who would find single attributes and sets of attributes
be well known to most undergraduates. that guaranteed that some object or person
The eighth proper name was the name of was an exemplar of a given word (i.e., suffi-
a fictitious person (Superman). The stim- cient attributes and sets of attributes). Sub-
uli can be found in the Appendix. jects were told that they might not be able
to find any necessary or any sufficient at-
Procedure tributes for some or even any of the words,
The experiment was conducted in four and that that was fine. In another session,
separate sessions. In the first session, subjects were asked to rate the importance
MODELS OF MEANING 453

o f each attribute for determining whether or subject. Similarly, a set o f attributes that a
not something was an exemplar o f a given subject rated as jointly sufficient and as in-
word. Subjects used a 1 - 7 scale for impor- dividually necessary was classified as a de-
tance ratings (1 = not at all important, fining set of attributes. Means were
7 = very important). Each subject's ratings 4.4 +_ 0.7 for natural kinds, 4.0 + 0.5 for
were made on his or her personal word-at- artifact terms, and 6.5 + 0.4 for proper
tribute list. (Procedural details for Experi- names, F(2,18) = 8.13, p < .005. Pairwise
ments 1 and 3 are reported in M c N a m a r a & comparisons o f the means (using Fisher's
Sternberg, N o t e 1). L S D test) showed that artifact terms and
natural kinds did not differ, but that proper
Results and Discussion
names differed significantly from both of
The mean numbers of words (out o f eight the others. This result is striking because it
possible) per subject with at least one nec- shows that the extent to which defining at-
essary attribute were 7.8 + 0.1 for artifact tributes could be identified for words de-
terms a n d 8.0 _+ 0.0 for both natural kinds pended on the type o f word. Examples o f
and p r o p e r names: E v e r y subject found at defining attributes are, for sandals, that
least one necessary attribute for e v e r y nat- they are shoes that are held on with straps
ural kind term and e v e r y p r o p e r name, and and do not c o v e r the whole foot; and for
all but one subject found at least one neces- diamond, that it is the hardest substance
sary attribute for each o f the artifact terms. known.
Examples o f some o f the necessary attri- To determine the extent to which sub-
butes are, for diamond, that it scratches jects perceived attributes to be necessary
glass, is the hardest substance known, as opposed to nonnecessary, the mean pro-
and is made of carbon; and for Albert Ein- portions o f necessary attributes per word
stein, that he is dead, was a scientist, was and subject were c o m p u t e d for each word
male, and that he invented the equation type. Means were .50 for natural kinds, .42
E = MC ~. for artifact terms, and .76 for proper names,
The mean numbers of words per subject F(2,18) = 49.30, p < .001. Pairwise com-
with at least one sufficient attribute or set parisons of the means showed that natural
of attributes were also quite high. Means kinds and artifact terms did not differ, but
were 8.0 ___ 0.0 for natural kinds, 7.6 ___ 0.2 that p r o p e r names differed from each of the
for artifact terms, and 7.5 --+ 0.2 for p r o p e r others. These means indicate that between
names. T h e m e a n s show that every subject 40 and 50% of the attributes listed for natu-
identified at least one sufficient set o f at- ral kinds and artifact terms were perceived
tributes for e v e r y natural kind term and that as necessary, on the average, whereas o v e r
subjects found sufficient sets o f attributes 75% of the attributes listed for p r o p e r
for just about all of the p r o p e r names and names were perceived as necessary, on the
artifact terms. Examples are, for an eagle, average.
that it is a bird that appears on quarters; Internal consistency reliabilities (coeffi-
and for a lamp, that it is a light source that cient ~) corresponding to splits on subjects
has a shade. were computed on those word-attribute
The most striking result is for the mean pairs that were seen by every subject.
numbers of words per subject with at least These 210 word-attribute pairs contained
one defining attribute or set o f attributes. only attributes on the master list and not id-
Recall that subjects n e v e r rated whether or iosyncratic ones. The reliabilities were .86
not an attribute or set o f attributes was both for the necessity ratings, .88 for the suffi-
necessary and sufficient. An attribute that a ciency ratings, and .81 for the importance
subject rated as sufficient and as necessary ratings. These reliabilities indicate that sub-
was classified as a defining attribute for that jects agreed to a large extent on which at-
454 MC NAMARA A N D STERNBERG

tributes were necessary and which were that confidence ratings would directly re-
sufficient or in a sufficient cluster. flect the weighted mean of the attributes in
Although we have shown that people the description.
perceive there to be necessary attributes
3. The mixed models. Two mixed models
and sufficient attributes, we do not know
were constructed by combining the defin-
whether people actually use these neces-
ing-attributes model with each of the addi-
sary or sufficient attributes when referring.
tive models.
The second experiment was conducted to
determine the relative degrees to which (a) The first mixed model combined de-
necessary, sufficient, and characteristic at- fining attributes with weighted sums of at-
tributes are used to refer, and more impor- tributes.
tantly, to test the three classes of models of (b) The second mixed model combined
meaning and reference. defining attributes with weighted means of
attributes.
EXPERIMENT 2
Both of these mixed models were of the
An identification task was used to test
kind discussed by Clark and Clark (1977,
versions of the three classes of models of
p.467): The representation of meaning con-
meaning and reference. This task required a
tains defining and characteristic attributes,
confidence rating that a described object,
or sometimes only characteristic attributes.
person, or concept was an exemplar of a
These five particular models are our spe-
word paired with the description. For ex-
cific instantiations of the three classes of
ample, subjects might see "Diamond"
models of meaning. They may or may not
paired with the description "Sparkles, Car-
agree with anyone else's exact specifica-
bon, Mined in Africa." Subjects would
tions. Our goals were to quantify the three
have to rate their confidence that this de-
classes of models of meaning in ways that
scribed object was in fact a diamond.
captured important aspects of the models
The following versions of the models
and that allowed them to be tested empiri-
were tested using this task.
cally.
1. A defining-attributes model. This
model predicted that if the description
Method
paired with the word was defining, then the
confidence rating on that item would be Subjects
high. However, if the description was not The subjects were nine out of ten sub-
defining, the model predicted that the confi- jects who participated in the first experi-
dence rating would be low. ment. One of the subjects in Experiment 1
2. The additive models. Two such char- was not available for Experiment 2. Sub-
acteristic-attribute models were formu- jects were compensated for their partici-
Fated. pation with course credit, monetary pay-
(a) A weighted-sum model predicted that ment, or both.
confidence ratings would directly reflect
the weighted sum of the attributes in the de- Materials
scription: higher weighted sums should In the following paragraphs, "descrip-
produce higher confidence ratings. A model tion" refers to a list of attributes. So,
in which identification depends only on the for example, a description for the word
number of matching attributes was viewed " p o t a t o " might be "Grows underground,
as a special case of the weighted-sum Brown skin, Cooked many ways." All de-
model, and was not separately tested. scriptions were constructed for a subject
(b) A weighted-mean model predicted using only those attributes on the sub-
M O D E L S OF M E A N I N G 455

ject's personal word-attribute list, which tion, (d) the weighted mean of attributes
had been compiled in the first experi- in a description, and (e) whether or not a
ment. Each description contained attri- description contained a negation. The
butes from one and only one word. weights for the attributes were obtained
Each subject's complete list of descrip- from subjects' importance ratings col-
tions contained the following items: (a) all lected in Experiment 1 on the 1-7 scale.
of the sets of attributes that the subject The weighted sum or weighted mean of a
rated both as necessary and as sufficient description was computed by taking the
in Experiment 1 (i.e, defining attributes sum or mean, respectively, of the impor-
and sets of attributes); (b) all of the de- tance ratings for the attributes in the de-
scriptions in (a) with one randomly se- scription. Since each subject saw a unique
lected attribute negated (e.g., "Grows list of descriptions, the weighted sums
underground" became "Does not grow and means of attributes were computed
underground"); (c) 144 new descriptions from a subject's own importance ratings
(6 per word) constructed from the sub- in Experiment 1. All independent vari-
ject's personal word-attribute list; and (d) ables were within subjects.
48 new descriptions (2 per word) that con- Dependent variable. The dependent vari-
tained one negated attribute. The descrip- able was confidence ratings provided in the
tions in (c) and (d) did not contain suffi- rating task.
cient attributes or sets of attributes (as
determined from subjects' ratings in Ex- Procedure
periment 1). The second experiment was conducted
The sizes of these complete lists ranged between two and three months after Ex-
from 230 to 306 items across subjects and periment 1. Subjects were given a rating
had a mean value of 260. Differences in form based on their personal description
the sizes of the description lists depended lists and were asked to rate their confi-
on how many defining sets of attributes dence that a described object or person
could be identified for the various sub- was an exemplar of the word paired with
jects. the description. The subjects were run in
A rating form was constructed for each groups of from one to three members and
subject from his or her complete list of the instructions were presented in written
descriptions. An item on this form con- form. Most subjects completed the task
sisted of a description paired with the ap- within one hour, but no time limit was im-
propriate word. Items were printed in a posed.
random order, and the order of presenta-
tion of attributes in an item was also ran- Results
domized. Each item was accompanied by The results will be divided into three
a labeled, 1 (DEFINITELY NO) to 8 sections: (a) the model fits for the three
(DEFINITELY YES) rating scale indicat- simple models; (b) the model fits for the
ing subjects' confidence that the describedtwo mixed models; and (c) analyses on
object was an exemplar of the word. descriptions containing negated attributes.
All group analyses were performed on
Design subjects' concatenated data sets, that is,
Independent variables. The independent there was no averaging across either sub-
variables were (a) whether or not a de- jects or items, and thus there was just one
scription was defining, (b) the number of observation per data point, for a total of
attributes in a description (1-18), (c) the 866 data points. Model tests were not per-
weighted sum of attributes in a descrip- formed on descriptions containing negated
456 MC N A M A R A A N D S T E R N B E R G

a t t r i b u t e s . 2 T h r e e o u t o f t h e 866 m o d e l e d t o o k t h e v a l u e 1 i f t h e a t t r i b u t e s in a d e -
items had missing confidence ratings and scription were defining for a given subject,
were eliminated from the analyses. The and 0 otherwise. For the weighted-sum
overall mean confidence rating on these model, the independent variable was the
863 i t e m s w a s 5.04 _+ 0.07. s u m o f t h e i m p o r t a n c e r a t i n g s f o r t h e attri-
b u t e s in a d e s c r i p t i o n . F o r t h e w e i g h t e d -
Model Fits for the Simple Models
mean model, the independent variable was
Each of the simple models contained one the mean of the importance ratings for the
independent variable. For the defining-at- a t t r i b u t e s in a d e s c r i p t i o n . A s a n e x a m p l e ,
tributes model, the independent variable consider the item "Yacht" paired with "A
b o a t , O w n e d b y w e a l t h y p e o p l e , H a s sails,
A l t h o u g h the d e s c r i p t i o n s c o n s t r u c t e d to be non-
sufficient (set c) could be as large as one less t h a n Large." This particular description was not
the number of attributes listed for a word, the sizes d e f i n i n g f o r t h e s u b j e c t f o r w h o m . it a p -
of the nonsufficient descriptions to be modeled were p e a r e d as a s t i m u l u s i t e m , so t h e v a l u e o f
limited. In Experiment 1, subjects were asked to find t h e d e f i n i n g - a t t r i b u t e s v a r i a b l e w a s 0. T h e
all attributes that by themselves were sufficient and
i m p o r t a n c e r a t i n g s p r o v i d e d b y this s a m e
to find clusters of attributes sufficient in conjunction.
When finding sufficient clusters, subjects were in- s u b j e c t in E x p e r i m e n t 1 w e r e 7 f o r " A
structed to find the smallest sets of attributes (in boat," 5 for "Owned by wealthy people," 6
terms of the number of attributes) that were suffi- for "Has sails," and 3 for "Large." Thus,
cient. Because subjects were required to find the the weighted-sum variable took the value
smallest possible sets, it is very likely, indeed almost
21 a n d t h e w e i g h t e d - m e a n v a r i a b l e t o o k t h e
certain, that there were larger sets of attributes that
subjects would have rated as sufficient had they v a l u e 5.25 (21/4). W e i g h t e d s u m s o f attri-
been so instructed. Hence, it is also very likely that biates r a n g e d f r o m 1 to 82 a c r o s s t h e 863
some of the nonsufficient descriptions constructed modeled items and had a median value of
for a word were sufficient because they contained 11. W e i g h t e d m e a n s o f a t t r i b u t e s r a n g e d
some number of attributes larger than the largest suf-
f r o m 1 to 7 a c r o s s t h e m o d e l e d i t e m s a n d
ficient cluster identified for that word. To provide a
test of the models that was fair to the defining-attri- h a d a m e d i a n v a l u e o f 5.2.
butes model, descriptions constructed for a word The three simple models were correlated
whose sizes exceeded the size of the largest suffi- w i t h c o n f i d e n c e r a t i n g s (a) o n t h e e n t i r e s e t
cient cluster identified for that word were not in- o f 863 i t e m s , (b) o n e a c h w o r d t y p e s e p a -
cluded in the modeled items. The size of the largest
r a t e l y , a n d (c) f o r e a c h s u b j e c t s e p a r a t e l y .
sufficient cluster identified for a word was, of
course, different for different subjects. If there were T h e o v e r a l l fits a n d t h e fits w i t h i n e a c h
no sufficient clusters identified for a word, then all w o r d t y p e , in t e r m s o f p r o p o r t i o n o f v a r i -
of the descriptions constructed for that word were a n c e a c c o u n t e d f o r (r2), a r e p r e s e n t e d in
included in the set modeled, regardless of their size. T a b l e 1. R o o t m e a n s q u a r e d e v i a t i o n is in-
This restricted set of items contained 559 out of the
appropriate because the predictions of the
1296 nonsufficient descriptions. The sizes of the de-
scriptions in this set ranged from 1 to 15 attributes models were on different scales (0-1 vs 1-
(mean = 3.1). The set of modeled items included 82 v s 1 - 7 ) .
these 559 items plus all of the necessary-and-suffi- It is c l e a r f r o m t h e p a t t e r n s o f m o d e l fits
cient descriptions, giving a total of 866 modeled in T a b l e 1 t h a t t h e b e s t p r e d i c t o r o f confi-
items across subjects. dence ratings was the defining-attributes
Descriptions with negated attributes were not
modeled because of unreasonable assumptions that model. This result held for the entire set of
must be made to model these items with additive 863 i t e m s a n d f o r e v e r y w o r d t y p e . B e c a u s e
models. For example, there is no reason to believe there were no replications of items across
that a mismatching attribute should be assigned the s u b j e c t s , it w a s i m p o s s i b l e to d e t e r m i n e
negation of the weight it gets when the attribute w h a t p r o p o r t i o n o f t h e reliable v a r i a n c e in
matches. We address this problem in Experiment 4
by collecting two kinds of importance ratings, one confidence ratings was accounted for by
for instances in which an attribute matches and an- the defining-attributes model. In any case,
other for instances in which an attribute mismatches. the defining-attributes model did quite well
M O D E L S OF M E A N I N G 457

TABLE I
MODEL FITS (r ~) FOR THE SIMPLE MODELS IN EXPERIMENT 2

Overall Natural kinds Artifact terms Proper names

Defining attributes .36*** .38*** .21"** .42***


Weighted sum .02* .03* .18"** .00
Weighted mean .11"** .04** .11"** .15"**
N u m b e r of data points 863 240 232 391

* p < .05.
** p < .01.
*** p < .001.

in accounting for between 21 and 42% of able was the sum of the importance ratings
the total variance in confidence ratings with
for all of the attributes in the description.
a single independent variable whose valuesThis variable was identical to the summed
were restricted to ones and zeroes only. importance index in the weighted-sum
The additive models made much less re- model. For the other mixed model (Model
stricted predictions, and so had the poten-
B), the second independent variable was
tial for accounting for more variance in the
the mean of the importance ratings for all of
ratings than the defining-attributes model.
the attributes in the description. This vari-
To verify that the patterns of model fits
able was identical to the mean importance
for the concatenated data were reflected for
index in the weighted-mean model.
individual subjects as well, the three simpleModel fits were computed for (a) the en-
models were correlated with confidence tire set of 863 items, (b) each word type
ratings for each subject separately. The de-
separately, and (c) each subject separately.
fining-attributes model correlated more Model fits were computed on subjects' con-
highly with confidence ratings than the two
catenated data sets. The overall fits and the
additive models for eight out of nine sub-fits within each word type (again, in terms
jects. The mean squared correlation for in-
of R 2) are presented in Table 2. Standard-
dividual subjects was .39 for the defining-
ized regression coefficients (/3 weights) for
attributes model, .07 for the weighted-sumthese two models are also presented in
model, and .15 for the weighted-mean Table 2. Because all parameters were pre-
model, F(2,16) = 23.87, p < .001. Pairwisedicted to be positive, significance tests for
comparisons of the means showed that the parameters are one-tailed.
weighted-sum and weighted-mean models The weighted-sum mixed model (Model
did not differ significantly, but that the de-
A) outperformed the weighted-mean mixed
fining-attributes model was significantly model (Model B) in every analysis. Model
better than each of the others. Thus, evenA accounted for between 36 and 51% of the
at the level of individual subjects, the defin-
total variance in confidence ratings with
ing-attributes model outperformed the only two independent variables (and on
other two simple models. data for which there was only one observa-
tion per modeled data point). The incre-
Model Fits for the Mixed Models mental contribution to R 2 provided by the
The mixed models shared one indepen- weighted-sum variable (over the r 2 for de-
dent variable, the necessity-sufficiency fining attributes alone) was substantial, .09,
index from the defining-attributes model. and highly significant, F(1,860)= 134.23,
The two mixed models differed with re- p < .001. It is possible that the success of
spect to how attributes were additively the weighted-sum mixed model is an arti-
weighted. For one of the mixed models fact created by (a) defining attributes
(Model A), the second independent vari- having very high importance ratings on av-
458 MC N A M A R A A N D S T E R N B E R G

TABLE 2
MODEL FITS (R 2) AND STANDARDIZED REGRESSION COEFFICIENTS (fl) FOR THE
TWO MIXED MODELS IN EXPERIMENT 2

Natural Artifact Proper


Overall kinds terms names

Defining attributes + weighted s u m


R~ .45*** .47*** .36*** .51"**
Standardized coefficients (/3)
Defining attributes .67*** .68*** .42*** .76***
Weighted s u m s .30*** .30*** .39*** .33***
Defining attributes + weighted m e a n
R2 .38*** .39*** .24*** .44***
Standardized coefficients (/3)
Def'ming attributes .56*** .60*** .39*** .59***
Weighted m e a n s .14"** .08 .18"* .16'**
N u m b e r o f data points 863 240 232 391

** p < .01 (significance tests for parameters are one-tailed).


*** p < .001.

erage, and (b) the existence of a curvilinear Descriptions Containing


relationship between weighted sums and Negated Attributes
confidence ratings (as one might expect Confidence ratings for descriptions with-
from a threshold model)? Although defin- out negated attributes were considerably
ing attributes had higher importance ratings higher than ratings for descriptions contain-
than nondefining attributes on average ing negated attributes. The mean rating for
(means of 5.7 and 4.9, respectively, on a 1- items without negations was 5.55-+ 0.05
7 scale), the difference was not large. More- (n = 1599) and the mean for items contain-
over, the scatterplot relating weighted sums ing negated attributes was 2.01 +-0.06
to confidence ratings did not reveal any (n = 727). The mean rating for affirmative
curvilinear trends. We concluded, there- necessary-and-sufficient descriptions was
fore, that the success of the mixed model 6.78 and the mean rating for the same
was not simply an artifact of the ways in descriptions containing a negated attri-
which variables enter linear regression. bute was 1.37. An analysis of variance on
The/3 weights show that the necessity- the mean ratings for defining descriptions
sufficiency index from the defining-attri- and their negated counterparts, computed
butes model contributed more to predicting for each subject separately, revealed a
subjects' confidence ratings than did the
highly significant effect of negation, F
additive index, and that this was true for
(1,8) = 703.59, p < .001.
both mixed models and for every kind of
We also compared ratings on items with
term. negated necessary attributes (excluding de-
When the two models were fit to each fining descriptions with negated attributes)
subject's data separately, Model A per- to ratings on items with negated nonneces-
formed as well as or better than Model B sary attributes. The mean confidence rating
for eight out of nine subjects. The mean for items with negated necessary attributes
squared multiple correlation for individual was 1.84 and the mean rating for items con-
subjects was .51 for Model A and .43 for taining negated nonnecessary attributes
Model B, F(1,8) = 8.73, p < .05. was 3.24. An analysis of variance on the
3 W e t h a n k an a n o n y m o u s reviewer for bringing mean ratings computed for each subject
this possible artifact to o u r attention. separately revealed that the difference
M O D E L S OF M E A N I N G 459

between the means was significant, F replicate the findings from Experiment 1 on
(1,8) = 28.20, p < .001. This result should a larger group of subjects.
be interpreted with some caution, however,
since the two kinds of descriptions varied Method
on a number of other variables (e.g., the
number of attributes) that w e r e not con- Subjects
trolled for. Subjects were 25 Yale undergraduates.
They were compensated for their partici-
pation with course credit and monetary
Discussion
payment.
The model tests in Experiment 2 showed
that a mixed model combining defining at- Materials
tributes and weighted sums of attributes
better predicted identification of a candi- The master list of words and attributes
date exemplar than (a) any of the simple constructed in Experiment 1 (see Appen-
models, and (b) a mixed model combining dix) was used in Experiment 3. Five ran-
defining attributes and weighted means of domizations of the words and attributes
attributes. These results were consistent on this list were used. Each sequence was
across subjects and word types. The suc- seen by five subjects.
cess o f this model indicates that the
methods of Experiment 1 (attribute listings Procedure
and ratings) measured aspects of the under- The experiment required from two to
lying representation of meaning: both the four sessions depending on how long it
necessity-sufficiency index and weighted took subjects to complete the rating tasks.
sums of attributes contributed significantly In three of the tasks, subjects made ne-
to the prediction of confidence ratings. Be- cessity, sufficiency, and importance rat-
cause subjects used attributes in ways that ings as in Experiment 1. Two new rating
directly reflected prior ratings on those at- tasks were introduced.
tributes, we are reasonably confident that The first new rating task was intro-
semantic attributes can be classified as de- duced to get a measure of an attribute's
fining or nondefining for a given subject, "decremental importance," so that ad-
and that attributes vary in their subjective ditive models could be used to model
importance. responses to items containing negated
The experiments so far tell us nothing attributes. Subjects were given a word-
about real-time processing of word mean- attribute list, but all of the attributes on
ings. The third and fourth experiments the list were negated. For example, for
were conducted to investigate both how the Paul Newman, " A c t o r " became " N o t an
meanings of nouns are mentally repre- actor" and " S e x symbol" became " N o t a
sented and how they are processed in real sex symbol." Subjects were asked to
time. Moreover, the new experiments en-
abled us to verify the results from the first Imagine that the only thing you know about
something is that the statement is true o f that
and second experiments on different sub- object or person. Given that that is all you
jects. know about something, I want you to rate your
confidence that it is N O T an example of the
EXPERIMENT 3 word. Suppose, for example, that one of the
The main reason for conducting the third words was " w h a l e " and one of the statements
was " d o e s not s w i m . " Given that you knew
experiment was to collect ratings needed that something did not swim, and that was all
for the fourth experiment. But in collecting you knew, I would want you to rate your con-
these ratings, it was possible to attempt to fidence that it was N O T a whale.
460 MC N A M A R A A N D S T E R N B E R G

Subjects used a 1-7 scale for their rat- Internal consistency reliabilities corre-
ings. sponding to splits on subjects were .88 for
The second new rating task was intro- standard importance ratings, .93 for decre-
duced to get a measure of an attribute's mental importance ratings, .98 for incre-
"incremental importance." Subjects were mental importance ratings, .93 for necessity
given the word-attribute list in its normal, ratings, and .93 for sufficiency ratings.
nonnegated form. The instructions re- These reliabilities are quite good and indi-
quested that subjects imagine, for each cate that the various subjects were per-
property listed below a word, " t hat the forming a given rating task in very similar
only thing you know about something is ways.
that it possesses that property. Given that
that is all you know . . . rate your confi- EXPERIMENT 4
dence that [the object or person] is an ex- Our goals for the last experiment were to
ample of the wo r d . " Subjects used a 1-7 verify the results of the second experiment
scale for their ratings. using converging operations and to test a
Results and Discussion model of real-time attribute information
processing. In particular, response latency
The findings from Experiment 1 were and response choice were used as depen-
replicated in Experiment 3. The mean dent measures, and the subjects' task was
numbers of words per subject with at to choose which of two described objects,
least one necessary attribute were people, or concepts was the better referent
7.7 --- 0.2 for natural kinds, 7.3 +- 0.3 for of a target word. For example, subjects
artifact terms, and 7.8---0.1 for proper might see " T e n t , " followed by two lists of
names. Although these means are slightly attributes: ( a ) " M a d e of canvas, Supported
lower than those found in Experiment 1, by poles, Portable, Waterproof," and (b)
the basic finding was the same: subjects " A shelter, Used for camping, Made of
found necessary attributes for just about canvas." Subjects would have to decide
all words used as stimuli. whether (a) or (b) was a better exemplar of
As in Experiment 1, subjects in Experi- a tent.
ment 3 found sufficient sets of attributes
for just about all of the words. The mean Method
numbers of words per subject with at
least one sufficient set of attributes were Subjects
7.9 --- 0.1 for natural kinds, 7.8 --_ 0.1 for The subjects were 21 of the 25 subjects
artifact terms, and 7.7---0.1 for proper who participated in Experiment 3, plus 11
names. new subjects who had not participated in
Finally, the mean numbers of words per any of the experiments. The four subjects
subject with at least one defining set of at- in Experiment 3 who did not participate
tributes were 4.9 --- 0.4 for natural kinds, in Experiment 4 either were unavailable
3.2 --- 0.4 for artifact terms, and 6.4 --- 0.4 when Experiment 4 was conducted or
for proper names. (Defining sets of attri- dropped out of the experiments. The sub-
butes were identified as in Experiment 1). jects were Yale University undergraduates
These differences were significant, F(2,48) and were compensated for their participa-
= 40.71, p < .001. Paired comparisons tion with course credit, monetary pay-
showed that all of the means were signifi- ment, or both.
cantly different from one another. Just as
in Experiment 1, necessary-and-sufficient Materials
sets of attributes could be identified for Experimental materials. The stimuli
more proper names than natural kinds or were 156 pairs of descriptions constructed
artifact terms. from attributes on the master list. De-
M O D E L S OF M E A N I N G 461

scriptions were constructed on the basis sary attributes in terms of the numbers of
of subjects' sufficiency and necessity rat- subjects for whom the attributes were
ings in Experiment 3. Necessary-and-suf- necessary.
ficient descriptions were not used (cf. Ex- Each subject saw a different random or-
periment 2) because we could not find dering of the 156 items. The order of pre-
enough unique sets of attributes that were sentation of the attributes in a description
both necessary and sufficient for large was also randomized for each subject.
numbers of subjects. The two possible orderings of the descrip-
The following kinds of items were in- tions within items were both used. Even
cluded as stimuli: (a) 48 pairs in which numbered subjects saw one ordering of
one description was "sufficient" and one the descriptions within the pairs and odd
description was "nonsufficient"; (b) 48 numbered subjects saw the opposite or-
pairs in which both descriptions were dering. Thus, each description was seen
"nonsufficient"; (c) 36 pairs in which one on the left about as many times as on the
description contained one negated "neces- right across subjects.
sary" attribute and the other description Reference ability tests. Subjects com-
contained one negated "nonnecessary" pleted two standardized ability tests: (a)
attribute; and (d) 24 pairs in which only the Nelson-Denny Reading Test-Form D
one description contained a negated at- (Brown, Nelson, & Denny, 1973), and (b)
tribute--for half of these pairs, the ne- the Verbal Reasoning subtest of the Dif-
gated attribute was "necessary" and for ferential Apptitude Tests-Form T (Ben-
the other half, the negated attribute was nett, Seashore, & Wesman, 1974).
"nonnecessary."
Because not all subjects agreed on Design
which attributes were sufficient, descrip- Independent variables. The independent
tions in (a) and (b) varied in terms of the variables were the (a) number of subjects
numbers of subjects for whom the de- in Experiment 3 for whom a description
scriptions were sufficient. The numbers of was sufficient, (b) number of subjects in
subjects for whom "sufficient" descrip- Experiment 3 for whom a negated attri-
tions were indeed sufficient ranged from bute was necessary, (c) number of attri-
10 to 25 (mean = 18.9, n = 48). The num- butes in a description (1-4), (d) weighted
bers of subjects for whom "nonsufficient" sum of attributes in a description, and (e)
descriptions were sufficient ranged from 0 weighted mean of attributes in a descrip-
to 7 (mean = 0.7, n = 144). Similarly, de- tion. The weights for the attributes were
scriptions in (c) and (d) varied in terms of obtained from subjects' mean incremental
the numbers of subjects for whom negated and decremental importance ratings col-
attributes were necessary. The numbers lected in Experiment 3. A nonnegated at-
of subjects for whom "necessary" attri- tribute was assigned a weight correspond-
butes were indeed necessary ranged from ing to that attribute's mean incremental
11 to 23 (mean = 19.3, n = 48). The num- importance rating. A negated attribute
bers of subjects for whom "nonneces- was assigned a weight corresponding to
sary" attributes were necessary ranged the negation of that attribute's mean dec-
from 0 to 9 (mean = 3.6, n = 48). Thus, remental importance rating. For example,
although there was some variation in the if the mean decremental importance rating
quality of the stimuli, the sufficient de- for an attribute was 6.7, then the weight
scriptions did not overlap with the nonsuf- assigned to that attribute, if it was ne-
ficient descriptions in terms of the num- gated in a description, would be -6.7. All
bers of subjects for whom the descriptions of the independent variables were within
were sufficient, and the necessary attri- subjects.
butes did not overlap with the nonneces- Dependent variables. The dependent
462 MC N A M A R A A N D STERNBERG

variables were (a) response latency, and groups were combined in subsequent anal-
(b) response choice. yses.
The internal consistency reliabilities
Procedure corresponding to splits on subjects were
The fourth experiment was conducted .97 for response choices and .90 for re-
between three and six months after Ex- sponse latencies. Response latencies out-
periment 3. Stimulus presentation and data side the "outer, upper fence" (Tukey,
collection were controlled by a NorthStar 1977) for each subject were classified as
microcomputer. Stimuli were presented on outliers and not included in the analyses.
a Televideo 920C terminal interfaced to the Outliers and observations missing due to
microcomputer. Subjects responded by machine failure accounted for less than
pressing one of two keys on the terminal 1% of the 4,992 observations. The overall
keyboard. mean response latency (with outliers re-
The target word was displayed for 1.5 moved) was 4.76 seconds (s = 3.06, n =
seconds on the terminal screen. One sec- 4962).
ond after the word was erased, the two
descriptions were displayed side by side Alternative Models of Meaning
on the video screen. Timing of response Representation: Response Choice
latencies began immediately after both de- Comparisons of alternative models of
scriptions had been displayed. Subjects meaning representation will be restricted
were instructed to decide which of the here to response choices. Fitting repre-
two described objects or people was the sentational models to response latencies
better exemplar of the target word. Sub- requires making processing assumptions,
jects were informed that there were no and so will be presented later.
right or wrong answers on the task, and
that they should take as much time as Quantification of the Models
they needed but no more to make their As in Experiment 2, three simple
decisions. Subjects responded by pressing models and two mixed models were
the " / " key if the described object on the tested. The simple models contained one
right was the better one and pressing the independent variable, which could take
" Z " key if the described object on the the value - 1 , 0, or 1. If a model pre-
left was the better one. Response choice dicted that the right description should be
and response latency were recorded on chosen over the left description, then the
every trial. The experiment lasted about variable took the value - 1. If a model did
30 minutes. not make a prediction (this was true only
In a later session, subjects completed of the defining-attributes model), then the
the ability tests. The ability testing lasted variable took the value 0. Finally, if a
one hour. model predicted that the left description
should be chosen over the right descrip-
Results tion, then the variable took the value 1.
Preliminary analyses showed that sub- The predictions of the defining-attributes
jects from Experiment 3 and new subjects model depended solely on whether a de-
agreed on 88% of their response choices. scription was sufficient or contained a ne-
In addition, mean response latencies on gated necessary attribute. For the 48
the items correlated .77 between the two items in which one description was suffi-
groups of subjects (.93 when this correla- cient, the defining-attributes model pre-
tion is corrected for attenuation). Because dicted that the sufficient description
of the good degree of correspondence be- would be chosen. Thus, this model took
tween the two groups, data from the two the value 1 or - 1 depending on whether
M O D E L S OF M E A N I N G 463

the sufficient description was on the left weighted sum or mean of attributes than
or the right, respectively. For the 48 items the left description.
in which neither description was suffi- The two mixed models tested shared
cient, the defining-attributes model took one independent variable, the one corre-
the value 0, indicating that it predicted sponding to the defining-attributes model.
each option equally. For the 48 items in The second variable was either the vari-
which one description contained a negated able corresponding to the weighted-sums
necessary attribute and the other con- model or the variable corresponding to
tained a negated nonnecessary attribute, the weighted-means model.
the defining-attributes model took the
value of 1 or - 1 depending on whether Tests of the Models
the description with the negated necessary The three simple models and two mixed
attribute was on the right or left, respec- models were fit to group response choices
tively. That is, when the negated neces- on the 156 items via linear regression.
sary attribute was on the right, the defin- The dependent variable measuring group
ing-attributes model predicted that the left choices took the value 1 or - 1 depend-
option would be chosen, and when the ing on whether the majority of subjects
negated necessary attribute was on the picked the left or right description, re-
left, the defining-attributes model predicted spectively. The subjects did not split
that the right option would be chosen. For fifty-fifty on any of the items. The mean
the 12 items in which one description con- percentage of subjects choosing the group
tained a negated necessary attribute and choice was 84%, indicating a high degree
the other contained no negations, the de- of consistency among subjects in response
fining-attributes model took the value 1 or choices.
- 1 depending on whether the negation The fits of the five models in terms of
was on the right or left, respectively. For proportions of variance accounted for (R 2)
the 12 items in which one description con- and root mean square deviation (RMSD)
tained a negated nonnecessary attribute, can be found in Table 3.
the defining-attributes model took the The results again showed that subjects
value 0. The weighted-sum and weighted- used necessary attributes and sufficient
mean models took the value 1 if the left attributes, and supported a mixed model
description had a greater weighted sum or combining defining attributes with weight-
mean of attributes than the right descrip- ed sums of attributes. (The models test-
tion. These models took the value - 1 ed in Experiment 4 were actually more
if the right description had a greater general than those tested in Experiment 2

TABLE 3
FITS OF THREE SIMPLE MODELS AND TWO MIXED MODELS TO
RESPONSE CHOICES IN EXPERIMENT 4

Root mean square Proportion of total


deviation variance accounted for

Defining attributes .73 .48***


Weighted sum .70 .57***
Weighted mean .80 .46***
Defining attributes
+ weighted sum .59 .65***
Defining attributes
+ weighted mean .66 .57***

*** p < .001.


464 MC N A M A R A A N D S T E R N B E R G

because the descriptions used as stimuli (1) Reading time was estimated by the
were not, in general, both necessary and total number of words in the two descrip-
sufficient.) The mixed model combining tions, excluding the target word. In the ex-
defining attributes and weighted sums of ample, the total number of words is 16. The
attributes outperformed all other models value of this variable ranged from 4 to 34
in terms of both RMSD and proportion of across items (mean = 14.0).
Variance accounted for. The incremental (2) Processing time for negations was
contribution to R 2 provided by defining at- estimated by the number of negated attri-
tributes, over the R z for weighted sums butes in the descriptions, which is 0 for the
of attributes alone, was .08, F(1,153)= example. The value of this variable ranged
34.27, p < .001. from 0 to 2 across items (mean = .6).
A Model for Processing of Meaning: (3) Time for comparison of attributes in
Response Latency the descriptions to attributes of the target
word was estimated by the total number of
Accepting a model of meaning repre- attributes in the two descriptions. Accord-
sentation in which nouns possess both ing to the model, each attribute in each
defining and characteristic attributes (or, description is compared to the attributes
in some cases, only characteristic attri- of the encoded target word. The weights
butes), we sought to test a model of how of matching and mismatching attributes
this attribute information is processed. are added to a weighted-sum counter for
This model assumed that (a) subjects the description currently being processed
tested both answer options to make sure (there is a weighted-sum counter for each
that they picked the better of the two, and description). Mismatching attributes are
(b) subjects compared answer options on also checked for necessity, and if they are
the basis of both defining attributes (when necessary, this information is recorded in a
present) and weighted sums of attributes. 4 defining-attributes counter for the descrip-
A flowchart for the model can be found in tion currently being processed (there is a
Figure 1. defining-attributes counter for each de-
Quantification of the Model scription). When all attributes in a descrip-
tion have been compared to the attributes
Quantification of the processing model
of the target word, the description is
will be explained by referring to the fol-
checked for sufficiency. If the description
lowing stimulus item: " T e n t " followed by
is sufficient, this fact is recorded in the de-
(a) "Made of canvas, Supported by poles,
fining-attributes counter for that descrip-
Portable, Waterproof," and (b) " A shel-
tion. In the example, the comparison vari-
ter, Used for camping, Made of canvas."
able would take the value 7, the number of
The six parameters of the model and the
attributes in the two descriptions. The com-
variables used to estimate them were as
parison variable ranged from 3 to 8 across
follows.
items (mean = 5.7).
4 In addition to the e x h a u s t i v e attribute informa- (4) Comparison of options on the basis
tion processing model, we also tested two "self-ter-
of defining attributes was estimated by the
m i n a t i n g " models. In t h e s e models, subjects were
p r o p o s e d s o m e t i m e s to c h o o s e an option only on the
absolute difference between the number of
basis o f def'ming attributes or only on the basis of subjects for whom the first description was
weighted s u m s of attributes. Fits for t h e s e m o d e l s sufficient and the number of subjects for
(R 2) were .78 for the m o d e l that serf-terminated on whom the second description was suffi-
weighted s u m s and .74 for the model that serf-termi- cient, or by the absolute difference between
n a t e d on defining attributes. B e c a u s e the differentia-
tion a m o n g the m o d e l s w a s n o t spectacular, we de-
the number of subjects for whom a negated
cided to report fits for the m o d e l with the highest R 2 attribute in the first description was neces-
a n d lowest R M S D , the e x h a u s t i v e model. sary and the number of subjects for whom
I Encode target word I

I Read descriptions I

Compare an a t t r i b u t e in one of the descriptions


I to a t t r i b u t e s of the target word

Add incremental weight to


weighted-sum counter
the current description
for

]Record t h i s fact in the


d e f i n i n g - a t t r i b u t e s counter
for the current description

Add decremental weight to


weighted-sum counter for
the current description

Y e s @

Record t h i s fact in the


d e f i n i n g - a t t r i b u t e s counter
for the current description

Y e s ~

Compare weighted-sum counters ]

Compare d e f i n i n g - a t t r ibutes counters [

~ [ Respond 1

I J u s t i f y choice on the basis


of de$i n i n g - a t t r i butes
]

IF~espondI
FIG.1.FlowchartfortheinformationprocessingmodeltestedinExperiment4.
465
466 MC N A M A R A A N D STERNBERG

a negated attribute in the second descrip- of definingness alone had to be justified.


tion was necessary. According to the The stimulus items were constructed so
model, comparison time decreases as the that for 60 of the 156 pairs of descriptions,
difference between the defining-attributes there were no differences in definingness
counter for the first description and the de- between the descriptions. For these items,
fining-attributes counter for the second de- the justifcation variable always took the
scription increases, that is, subjects are value 0, since there could be no discrep-
faster the more dissimilar the options are. ancy between choices. The difference in
We needed to use a continuous variable to weighted sums and the difference in defn-
estimate a dichotomous construct (the ne- ingness predicted opposite choices for 10 of
cessity or sufficiency of a set of attributes) the remaining 96 items. For these items, the
because we were modeling group average justification variable took the value 1. In
data and a given description was not the example, both the difference in defin-
equally good or bad for all subjects. The ingness and the difference in weighted sums
difference between the two descriptions predict that the second option should be
capitalized on this inherent variability in chosen. Thus, the value of the justification
our stimuli. This comparison variable was variable would be 0.
linearly scaled so that small values on the
variable corresponded to large differences Tests of the Model
between the two descriptions, and hence to The model described above was tested in
fast comparison times. In the example, the terms of its ability to account for mean re-
first description was sufficient for none of sponse latencies on the 156 items. Re-
the subjects and the second was sufficient sponses were included in the mean laten-
for 11 subjects. Thus, the comparison vari- cies even if they were "errors" under the
able was a linear function of the number 11 model. Fits changed trivially when errors
(precisely 26-11, or 15). were excluded. Fits of the model, along
(5) Comparison o f options on the basis with standardized regression coefficients/3
of weighted sums of attributes was esti- for the parameters, can be found in Table 4.
mated by the absolute difference in
summed weights between the two descrip- TABLE 4
tions. It was assumed that comparison time FIT OF THE PROCESSING MODEL TO RESPONSE
decreases as the difference between the LATENCIES IN EXPERIMENT 4
weighted-sum counter for the first descrip- R2 .79***
tion and the weighted-sum counter for the RMSD .66
second description increases. This vari- F Regression 91.88
able, like (4), was linearly scaled so that Standardized regression
small values on the variable corresponded coefficients (/3)
to large differences between the two de-
Reading time .42***
scriptions. In the example above, the first Processing
description had a weighted sum of 11.32 negations .37 ***
and the second description had a weighted Comparison to
sum of 11.44. Hence, the comparison vari- target w o r d .33 **~
Definingness
able was a linear function o f . 12 (precisely
comparison .23***
17.76-.12, or 17.64). Weighted-sum
(6) The justification parameter was rele- comparison .33***
vant when the difference in summed Justification .07*
weights and the difference in defining at- * p < .05 (significance tests for parameters are
tributes predicted opposite choices. In such one-tailed).
cases, the choice of an option on the basis *** p < .001.
MODELS OF MEANING 467

The proposed model provides a good ac- cremental contribution to the regression
count of the processing of attribute infor- equation.
mation; it accounts for nearly 80% of the
total variance in response latencies (and
with only six independent variables on 156 Discussion
data points). The standardized regression The results from Experiment 4 converge
coefficients for the model seem generally nicely on the results from Experiment 2.
reasonable. The /3 weights indicate that The analyses on response choices and re-
weighted sums of attributes were somewhat sponse latencies indicated that subjects
more important than defining attributes in used two kinds of information to compare
deciding which option was the better exem- exemplars of the nouns: (a) whether de-
plar of the target word. scriptions were sufficient or were missing
necessary attributes, and (b) weighted sums
Correlations with Ability Tests of attributes. The second major finding in
Experiment 4 was our success in modeling
Correlations were computed between the processes used to make these fine judg-
overall mean latencies on the decision task, ments of meaning. We believe that these
Vocabulary, Comprehension, and Read- two sets of findings, together with the re-
ing Rate scores from the Nelson-Denny, sults of Experiment 2, provide convincing
and Verbal Reasoning scores from the Dif- evidence for a mixed model of meaning and
ferential Aptitude Tests. These correlations reference.
were based on 31 of the 32 subjects who The correlation between latency and a
participated in Experiment 4. One subject combined index of reading rate and com-
was eliminated because his ability test prehension suggests that processes used in
scores were clear outliers (i.e., were ex- our decision task may overlap with pro-
tremely low) and unduly affected the pat- cesses used in certain verbal comprehen-
terns of correlations. sion problems. We have argued elsewhere
The only significant correlation involving (Sternberg & McNamara, in press) that
latency was that between overall mean la- processing models similar to the one tested
tency and Reading Rate, r - - - . 3 7 , p < in Experiment 4 could be used to model the
.05. However, the multiple correlation be- real-time processing of synonym problems.
tween mean latency, on the one hand, and In these problems, individuals must often
both Reading Rate and Comprehension, on discriminate fine shades of meaning, much
the other, was .47, F(2,28) = 4.04,p < .05. as in our task, to choose the correct option.
(The correlation between Comprehension In the future, we plan to test attribute infor-
and Reading Rate was .07 and the correla- mation processing models using tasks very
tion between Comprehension and mean la- similar to standard synonym problems.
tency was .27. These correlations were not
significant.) Both Reading Rate and Com-
prehension made statistically significant G E N E R A L DISCUSSION
contributions to the multiple correlation The results from the four experiments
(/3 = .38 for rate and .30 for comprehen- taken together provide a reasonably coher-
sion). Thus, Reading Rate and Comprehen- ent picture of both the representation of
sion, when taken together, were moder- meaning and the processes used to refer.
ately strongly related to latency in our task. First, with respect to the representation of
This relationship did not selectively involve meaning, subjects in the first and third ex-
the reading time parameter of the informa- periments found necessary attributes and
tion processing model. Vocabulary score sufficient attributes for nearly all of the
did not make a statistically significant in- words used as stimuli. Subjects' ratings of
468 MC N A M A R A A N D S T E R N B E R G

necessity and sufficiency allowed identifi- Related Models


cation of necessary-and-sufficient (defin-
ing) attributes for about 50% of the natural The weighted-sum mixed model quanti-
kinds and artifact terms, on average, and fies and extends a mixed model discussed
nearly 80% of the proper names, on aver- by Clark and Clark (1977, p.467). However,
age. (These defining attributes were not de- there are two important differences be-
fining for all of our subjects. We were con- tween the two models. First, the model
cerned with which sets of attributes were tested here combines attributes as a
necessary-and-sufficient for a given sub- weighted sum rather than as an unweighted
ject, not for all subjects in the experi- sum (the number of attributes). Second, in
ments.) That these percentages are below the weighted-sum mixed model, referring
100% is not surprising, since attempts to depends on both defining attributes (when
identify defining attributes of at least some they are present) and weighted sums of all
words have often been unsuccessful (e.g., attributes (both defining and characteris-
Hampton, 1979, 1981; Wittgenstein, 1953). tic). In the mixed model outlined by Clark
However, in another respect, these results and Clark, referring depends on defining at-
are quite novel in that a number of words tributes or an unweighted sum of attributes,
did have defining attributes, and in that but not both. We did not test models based
there seemed to be systematic differences on unweighted sums of attributes in our ex-
across word domains in the frequency of periments. However, our data do suggest
words with defining attributes. Results that two kinds of information are used to
from the first and third experiments tended identify and compare exemplars of con-
to support a mixed representation, in which cepts.
the meanings of nouns contain defining at- Another model that is similar in some
tributes and characteristic attributes, or in ways to a weighted-sum mixed model has
some cases, only characteristic attributes. been proposed by Smith and Medin (1981).
This conclusion is at best tentative without In the "probabilistic feature model," the
corroboration from the second and fourth meanings of concepts are given by sets of
experiments, since Experiments 1 and 3 say features that are weighted according to
nothing about the extent to which people their combined salience and conditional
use defining and characteristic attributes. probability. For example, the feature
Mere identification of defining attributes is "winged" for the concept "bird" would
poor evidence in support of a mixed model have a high weight since it is salient and al-
of meaning, if subjects never use those at- ways occurs with members of the concept,
tributes. whereas the feature "sings" would have a
The second and fourth experiments indi- lower weight since it is less likely to occur
cated that subjects used necessary, suffi- with birds (although it is fairly salient).
cient, and characteristic attributes. This re- These weights are on a 0-1 scale, with the
sult lends further support for a mixed upper limit representing features of high sa-
semantic representation. These two experi- lience and with conditional probability of 1,
ments also showed that a mixed model of that is, necessary (but nonsufficient) fea-
referring, in which identification of a candi- tures. The probabilistic feature model also
date exemplar depends on both defining at- allows sufficient but nonnecessary features
tributes and weighted sums of all attributes, to be included in the representation of
best predicted subjects' responses in a rat- meaning. In addition to their representa-
ing task and a timed decision task. It should tional assumptions, Smith and Medin make
be emphasized that this model makes no one general processing assumption, viz,
claim about the universality of particular at- that a candidate exemplar is categorized as
tributes; different people might have some- an instance of a concept if the candidate
what different sets of defining or character- possesses some critical sum of weighted
istic attributes. features of the concept.
MODELS OF MEANING 469

The similarities between the probabilistic and empirical evidence against traditional
feature model and the weighted-sum mixed theories of meaning. The authors show that
model are clear. Both models propose that a probabilistic feature model readily han-
the meaning of a concept is represented as a dles all of these problems. A weighted-sum
set of attributes weighted according to their mixed model can also easily handle these
importance. (In the following discussions, seven criticisms of traditional theories. (Be-
we will use "attribute" and "feature" in- cause a mixed model and the probabilistic
terchangeably. Note, though, that "fea- feature model deal with these criticisms in
ture" is quite constrained in meaning for very similar ways, we will not discuss them
Smith and Medin, much more so than "at- in any detail.) For example, disjunctive
tribute" is for us.) Second, both models concepts are not a problem for mixed
propose that categorization depends on the models because not all words have neces-
weighted sum of attributes that match be- sary or defining attributes under this view.
tween an object, person, or concept and the Moreover, since the same weighted sum of
intension of the concept. attributes can be achieved by various com-
The models are quite different, however, binations of matching and mismatching at-
in other ways. tributes, several different sets of attributes
First, the weighted-sum mixed model, could be used to determine category mem-
unlike the probabilistic feature model, bership (Smith & Medin, 1981). To account
makes fairly specific claims about the kinds for typicality effects, it is assumed (just
of attributes that are included in the repre- as in the probabilistic feature model)
sentation of meaning. In particular, the that judged typicality directly reflects the
mixed model proposes that some words weighted sum of attributes that match and
have defining attributes and that these at- mismatch between an exemplar and a con-
tributes are used in categorization. Results cept. Typicality does not depend on defin-
from our experiments support this pro- ing attributes except insofar as they enter
posal. into the computation of the weighted sum.
Second, the weighted-sum mixed model, So, for example, rabbits would be rated as
unlike the probabilistic feature model, more typical of mammals than whales
makes quite detailed processing claims (see would, because rabbits share more (or
Fig. 1). For example, the weighted-sum more highly weighted) attributes with most
mixed model that we tested proposes that other mammals (and hence, with the con-
individuals use two kinds of information to cept "mammal") than do whales. That
make category decisions (when both kinds "rabbit" shares more (or more highly
of information are indeed available). An- weighted) attributes with "mammal" than
other processing claim made by the mixed " w h a l e " can also explain typicality effects
model is that the weights of both matching on categorization times. Presumably, the
and mismatching attributes are entered into weighted-sum counter would reach its
the weighted sum, thus providing a mech- membership criterion faster for typical
anism for deciding when something is not members than for atypical members be-
an exemplar of a concept. The probabilistic cause there are more matching attributes
feature model lacks such a mechanism. and fewer mismatching attributes for typi-
All in all, we think that the weighted-sum cal than for atypical members. Defining at-
mixed model represents a significant ad- tributes (if present) would have no effect on
vance over the probabilistic feature model categorization times because they would be
because it is more restricted in terms of the the same for all exemplars. (See Smith and
representations and processes it proposes, Medin (1981) for a discussion of how simi-
and hence, is more falsifiable. lar logic explains results on nested triples of
In their review of the literature on concepts (e.g., chicken-bird-animal) and
conceptual representation, Smith and Me- on determinants of typicality (Rosch &
din (1981) identify seven lines of logical Mervis, 1975)).
470 MC N A M A R A A N D S T E R N B E R G

All of the models tested in our experi- ing or characteristic attributes. Moreover,
ments assumed that (a) word meanings can we examined only 24 different nouns. Goals
be decomposed into semantic attributes, of future research should be to determine
and (b) these attributes are used to refer. for what other kinds of words sets of attri-
However, both of these assumptions are butes provide a reasonable account of the
now under fire from psychologists, philoso- mental representation of meaning, and to
phers, and linguists. First, some psycholo- generalize the present research to a larger
gists and linguists have argued against the set of nouns.
view that word meanings are given by se- Second, it might be argued that subjects
mantic markers or attributes. For example, identified necessary attributes and suffi-
in a recent article, Fodor, Garrett, Walker, cient attributes for so many words because
and Parkes (1980) present both logical and of demands imposed by the tasks. This is a
empirical evidence against semantic decom- reasonable criticism of our results (and of
position of word meanings. Second, some rating tasks in general), and one that cannot
philosophers (e.g., Kripke, 1972; Putnam, be explicitly tested. Subjects were told that
1977) have argued that feature or attribute some or all of the words might not have
lists are an insufficient account of word necessary or sufficient attributes, but this
meanings, especially for proper names. certainly does not guarantee that subjects
They have argued that proper names, might not feel that we expected them to find
and possibly natural kind terms, refer these kinds of attributes anyway. Surely,
independently of identifying descriptions. though, task demands could not account for
The arguments used to defend these two the magnitude of the results. Even if half
attacks on "definitional" theories of mean- of the effects were due to task demands
ing are much beyond the scope of this arti- (which seems unlikely), subjects neverthe-
cle. We consider both of these criticisms, less would have found necessary attributes
and evidence on definitional theories, else- and sufficient attributes for about 50% of
where (McNamara & Sternberg, Note 2). the words on average. This percentage, in
Our investigations have led us to believe itself, would have been sizable.
that the evidence against definitional the- Third, the present series of experiments
ories is less compelling than some have uses a restricted sample of tasks. The two
argued (see, especially, Armstrong, Gleit- tasks used in Experiments 2 and 4 were
man, & Gleitman, in press; Fodor et al., fairly different in the kinds of demands
1980). made of the subjects, yet the tasks were
similar in other ways, for example, the at-
Limitations on the Experiments tributes needed to make the appropriate de-
The methods and results of the four ex- cisions were always explicitly available.
periments reported here are limited in sev- The similarities of the tasks may have con-
eral ways. tributed to the consistency of the results
First, the sample of words used as stimuli across the two tasks. Although we would
is fairly restricted. The consistency of our argue that a weighted-sum mixed model
findings across the four experiments sug- best describes the representations and pro-
gests to us that meanings of most nouns cesses used in our tasks, we do not know
could be given by sets of defining and char- that such a model is the best model in all
acteristic attributes (or sometimes only tasks. In fact, the representations and pro-
characteristic attributes). However, it is cesses used to make meaning judgments
not at all clear that the meanings of words may differ from task to task. A goal of
other than nouns (e.g., verbs) could be future research should be to assess the
given by sets of attributes, let alone defin- generalizability of a weighted-sum mixed
MODELS OF MEANING 471

model to tasks different from the ones used vergent validation from tasks, such as rec-
in our experiments. Synonym problems ognition priming (e.g., Ratcliff & McKoon,
(see discussion of Experiment 4) might be 1978), that control strategic variables more
ideal for testing models of meaning: first, fully than did the tasks that we employed in
semantic attributes are surely used in these our experiments.
problems, if attributes are used at all; and
second, semantic attributes are not explic-
itly available as they were in the tasks that Conclusion
we used. Despite the limitations on the experi-
A fourth limitation of our experiments is ments, several conclusions can be drawn.
the restricted sample of subjects. Keil and First, the results indicate that semantic at-
Batterman (Note 3) have shown that there tributes of at least some nouns can be clas-
are systematic differences across ages in sifted as defining or characteristic for a given
the ways that people represent concepts. individual. Second, the role of defining at-
Their data suggest that young children's tributes is more pronounced for some nouns
representations of concepts tend to consist (e.g., proper names) than for others (e.g.,
of characteristic attributes, rather than de- natural kind terms). Third, and finally, when
fining ones, whereas older children's repre- identifying and comparing the exemplars of
sentations tend to contain defining attri- certain concepts, individuals seem to use
butes as well as characteristic attributes. both defining attributes (when present) and
Our results are consistent with Keil and weighted sums of all attributes.
Batterman's, insofar as the subjects in our
experiments (adults) seemed to use both
defining and characteristic attributes. It
A P P E N D I X : M A S T E R L I S T OF WORDS AND
may very well be, however, that a purely
additive model might better account for the ATTRIBUTES C O M P I L E D IN E X P E R I M E N T 1
representation and processing of word
meanings among children. Production
A fifth limitation of our experiments con- frequency
Natural kind terms
cerns the difficulty in specifying the under-
1. Potato
lying representation of meaning. We be- (a) Grows underground 7
lieve that the consistency of our results (b) Brown skin 7
across tasks and subjects lends support to (c) Cooked many ways 7
the contention that the underlying repre- (d) A root 6
sentation of meaning contains defining and (e) A vegetable 5
(f) Edible 4
characteristic attributes weighted accord- (g) Contains starch 4
ing to their subjective importance. How- (h) White inside 3
ever, it is possible that the attributes listed (i) Has buds called " e y e s " 3
by our subjects were constructed at the (j) Round shape 2
time of task performance from a more basic 2. Diamond
(a) Valuable 10
representation that little resembles listed (b) Hardest substance known 9
attributes. It is also possible that the neces- (c) Rare 7
sity, sufficiency, and importance ratings (d) Crystalline 4
did not accurately measure properties, as it (e) Sparkles 4
were, of the underlying representation of (f) Carbon 3
(g) Mined in Africa 3
meaning. Specification of the underlying (h) Used in jewelry 3
representation (if there is in fact a single un- (i) Scratches glass 2
derlying representation) will require con- (j) Reflects light 2
472 MC NAMARA AND STERNBERG

3. Tiger (f) Curved 3


(a) Member of the cat family 7 (g) Peel to eat 3
(b) Carnivorous 6 (h) Yellow skin 3
(c) Large 6 (i) Grows in bunches 2
(d) Lives in Asia 6 (j) White inside 2
(e) An animal 5 Artifact terms
(f) Dangerous 5 1. Yacht
(g) Black stripes 3 (a) A boat 9
(h) Orange stripes 3 (b) Expensive 8
(i) Quick 3 (c) Large 6
(j) Strong 3 (d) Has sails 4
4. Eagle (e) Has a motor 3
(a) A bird 10 (f) Luxurious 3
(b) Symbol for the U.S.A. 9 (g) Owned by wealthy people 3
(c) Endangered species 7 (h) Associated with wealth 3
(d) Predator 6 (i) Used for leisure 2
(e) Large 4 2. Happiness
(f) Flies 4 (a) Good feeling 3
(g) Beautiful 3 (b) Something people strive for 2
(h) Looks mean 3 (c) Feeling good 2
(i) Builds nests 2 (d) State of joy 2
5. Maple (e) State of contentment 2
(a) Has sweet sap that is used to (f) Causes one to smile 2
make syrup 8 3. Lamp
(b) A tree 7 (a) Light source 10
(c) Loses its leaves in the winter 6 (b) Has a bulb 5
(d) Has leaves 5 (c) Electrical 5
(e) Its leaf is on the Canadian flag 3 (d) Placed on tables 3
(f) Leaves have several points 3 (e) Produces heat 2
(g) A pail hanging on it is used to (f) Has a shade 2
collect the sap 3 (g) Household object 2
(h) Common in New England 2 (h) Placed on the floor 2
6. Copper (i) Used for decoration 2
(a) A metal 9 4. Scientist
(b) Good conductor 6 (a) Tries to understand the world 5
(c) Malleable 6 (b) Engages in research 3
(d) Used to make pennies 5 (c) Intelligent 3
(e) Used to make wire 5 (d) Educated 3
(f) Bright 4 (e) Performs experiments 2
(g) An element 3 (f) Inquisitive 2
(h) Valuable 3 (g) Formulates hypotheses 2
(i) Used in jewelry 3 (h) Tries to prove or disprove hy-
7. Salmon potheses 2
(a) A fish 9 (i) Makes discoveries 2
(b) Edible 8 (j) Works in a laboratory 2
(c) Swims upstream to spawn 7 5. Violin
(d) Lives in saltwater 5 (a) Has four strings 11
(e) Has red flesh 3 (b) Musical instrument 10
(f) Large 2 (c) Played with a bow 7
(g) Dies after spawning 2 (d) Played in orchestras 3
(h) Preserved by smoking 2 (e) Played in classical music 3
(i) Preserved by canning 2 (f) Made of wood 3
(j) Silver colored 2 (g) Smallest instrument of its kind 2
8. Banana (h) Produces a high-pitched sound 2
(a) A fruit 9 6. Sandals
(b) Tropical 6 (a) Worn in warm weather 10
(c) Monkeys like to eat them 5 (b) A shoe 9
(d) Sweet 4 (c) Open-toed 6
(e) Soft 4 (d) Made of leather 6
MODELS OF MEANING 473

(e) Held on with straps 4 5. Moses


(f) Does not cover the whole foot 3 (a) Biblical figure 9
(g) Made of rubber 3 (b) Received the Ten Command-
(h) Comfortable 2 ments 7
(i) Made of straw 2 (c) Leader of his people 4
7. Wisdom (d) Parted the Red Sea 4
(a) Acquired with age 3 (e) Jewish 3
(b) Acquired with experience 3 (f) Male 3
(c) Knowledge 2 (g) Old 3
(d) Attribute of a person 2 (h) Led his people out of Egypt 3
(e) Good thing to have 2 (i) Wise 2
8. Tent (j) A prophet 2
(a) A shelter 10 6. Shakespeare
(b) Used for camping 9 (a) Playwright 11
(c) Made of canvas 4 (b) English 7
(d) Supported by poles 3 (c) Famous 5
(e) Anchored to the ground with (d) Poet 4
stakes 3 (e) Wrote many sonnets 3
(f) Portable 3 (f) Wrote many plays 2
(g) Waterproof 2 (g) Lived in the 16th Century 2
(h) Green 2 (h) Great 2
Proper names (i) Wrote "Romeo and Juliet" 2
1. Queen Elizabeth I1 7. Paul Newman
(a) Queen of England 10 (a) Actor 11
(b) A woman 3 (b) Acts primarily in films 5
(c) Not powerful 3 (c) Races automobiles 4
(d) Wealthy 2 (d) In "Butch Cassidy and The Sun-
(e) Famous 2 dance Kid" 4
(f) Member of nobility 2 (e) Blue eyes 3
(g) Her position is primarily for (f) Famous 3
show 2 (g) Handsome 3
2. Albert Einstein (h) Male 2
(a) Physicist 7 (i) Sex symbol 2
(b) Genius 7 (j) In "The Sting" 2
(c) Devised the theory of relativity 5 8. Superman
(d) German 2 (a) Comic book character 9
(e) Famous 2 (b) Flies 8
(f) Invented the equation (c) Super strong 8
" E = MC ~'' 2 (d) From the planet Krypton 4
(g) Male 2 (e) Harmed by Kryptonite 4
(h) Dead 2 (f) Fights evil 4
(i) Mathematician 2 (g) Disguised as Clark Kent 4
(j) White hair 2 (h) Invincible 4
3. Aristotle (i) Faster than a speeding bullet 2
(a) Philosopher 10 (j) Fights for justice 2
(b) Greek 9
(c) Lived long ago 2
REFERENCES
(d) Student of Plato 2
(e) Famous 2 ANDERSON, N. H. Algebraic rules in psychological
(f) Formalized logic 2 measurement. American Scientist, 1979, 67,
4. Richard Nixon 555-563.
(a) Expresident of the U.S.A. 9 ARMSTRONG, S. L., GLEITMAN, L. R., • GLEIT-
(b) Resigned from office 6 MAN, H. On what some concepts might not be.
(c) Almost impeached 3 Cognition, in press.
(d) Lives in California 3 BATTIG, W. F., & MONTAGUE, W. E. Category
(e) Unpopular 3 norms for verbal items in 56 categories: A repli-
(f) Wealthy 2 cation and extension of the Connecticut Cate-
(g) A crook 2 gory Norms. Journal of Experimental Psychol-
(h) Male 2 ogy Monograph, 1969, 80, (3, Pt. 2).
474 MC NAMARA AND STERNBERG

BENNETT, G. K., SEASHORE, H. G., & WESMAN, mantic information. Journal of Experimental
A. G. Differential Aptitude Tests-Form T. New Psychology: Learning, Memory, and Cognition,
York: The Psychological Corporation, 1974. 1982, 8, 16-36.
BROWN, J. I., NELSON, M. J., & DENNY, E. C. Nel- RoscH, E. On the internal structure of perceptual
son-Denny Reading Test-Form D. Boston: and semantic categories. In T. E. Moore (Ed.),
Houghton Mifflin, 1973. Cognitive development and the acquisition of
CLARK, H. H., & CLARK, E. V. Psychology and lan- language. New York: Academic Press, 1973.
guage. New York: Harcourt Brace Jovanovich, RoscH, E., & MERVIS, C. B. Family resemblances:
1977. Studies in the internal structure of categories.
FODOR, J. A., GARRETT, M. F., WALKER, E. C. T., Cognitive Psychology, 1975, 7, 573-605.
& PARKES, C. H. Against definitions. Cognition, RUSSELL, B. On denoting. In R. C. Marsh (Ed.),
1980, 8, 263-367. Logic and knowledge. London: Allen and
FREGE, G. On sense and reference. In P. Geach & Unwin, 1956.
M. Black (Eds.), Translations from the philo- SCHWARTZ, S. P. Naming, necessity, and natural
sophical writings of Gottolb Frege. Oxford: kinds. London: Cornell Univ. Press, 1977.
Basil Blackwell & Mott, 1952. SMITH, E. E., & MEDIN, D. L. Categories and con-
GLASS, A. L., & HOLYOAK, K. J. Alternative con- cepts. Cambridge: Harvard Univ. Press, 1981.
ceptions of semantic memory. Cognition, SMITH, E. E., SHOBEN, E. J., & RIPS, L. J. Struc-
1974/1975, 3, 313-339. ture and process in semantic memory: A fea-
HAMPTON, J. A. Polymorphous concepts in semantic tural model for semantic decisions. Psychologi-
memory. Journal of Verbal Learning and Verbal cal Review, 1974, 81, 214-241.
Behavior, 1979, 18, 441-461. STERNBERG, R. J., • MCNAMARA, T. P. The repre-
HAMPTON, J. A. An investigation of the nature of sentation and processing of information in real-
abstract concepts. Memory & Cognition, 1981, time verbal comprehension. In S. E. Whitely
9, 149-156. (Ed.), Test design: Contributions .from psychol-
KATZ, J. J. Semantic theory. New York: Harper & ogy, education, and psychometrics. New York:
Row, 1972. Academic Press, in press.
KRIPKE, S. A. Naming and necessity. In D. David- TUKEY, J. W. Exploratory data analysis. Reading,
son & G. Harman (Eds.), Semantics of natural Mass.: Addison-Wesley, 1977.
language. Dordrecht: Reidel, 1972. WITTGENSTEIN, L. Philosophical investigations. Ox-
KUCERA, H., & FRANCIS, W. H. Computational ford: Basil Blackwell & Mott, 1953.
analysis of present-day American English. Prov-
idence: Brown Univ. Press, 1967.
REFERENCE NOTES
McCLOSKEY, M., & GLUCKSBERG, S. Decision pro-
cesses in verifying category membership state- 1. MCNAMARA, T. P., & STERNBERG, R. J. The rep-
ments: Implications for models of semantic
resentation of meaning and mechanisms of ref-
memory. Cognitive Psychology, 1979, 11, 1-37. erence. Unpublished manuscript, 1982.
PUTNAM, H. Meaning and reference. In S. P.
2. MCNAMARA, T. P., & STERNBERG, R. J. Attri-
Schwartz (Ed.), Naming, necessity, and natural
butes of theories of meaning. Manuscript sub-
kinds. London: Cornell Univ. Press, 1977. mitted for publication, 1982.
RATCLIFF, R., t~ McKooN, G. Priming in item rec-
3. KEIL, F. C., & BATTERMAN, N. A characteristic-
ognition: Evidence for the propositional struc-
to-defining shift in the development of word
ture of sentences. Journal of Verbal Learning meaning. Manuscript submitted for publication,
and Verbal Behavior, 1978, 17, 403-417. 1982.
RATCLIFF, R., d?z McKooN, G. Speed and accuracy
in the processing of false statements about se- (Received June 9, 1982)

You might also like