Cognitive Psychology and Its Implications, Ch. 5
5Representation of Knowledge
Recall a wedding you attended a while ago. Presumably, you can remember who
married whom, where the wedding was, many of the people who attended, and
some of the things that happened. You would probably be hard pressed, however,
to say exactly what all the participants wore, the exact words that were spoken,
the the way the bride walked down the aisle, and so on—although you probably
registered many of these details. It is not surprising that after time has passed, our
memories lose some of the information of the original experience. What is interesting
is that our loss of information is selective: We tend to remember the gist
(that which is most meaningful or useful) and forget the detail (that which is not
important).
The previous chapter was about our ability to form visual images of the detail of
our experiences. It might seem that it would be ideal if we had the capacity to always
represent information in such detail and remember the detail. However, the histories
of the few individuals who have such detailed memories suggest that this might be a
curse rather than a blessing. Luria (1987) describes the story of a Russian journalist
who lived in the first half of the 20th century, who had very vivid imagery and the
ability to remember a great many perceptual details of his experience. He had problems
with many aspects of ordinary life, including reading:
As he put it: “Other people think as they read, but I see it all.” As soon as he began
a phrase, images would appear; as he read further, still more images were evoked,
and so on. . . . If a passage were read to him quickly, one image would collide with
another in his mind; images would begin to crowd in upon one another and would
become contorted. (p. 112)
Such problems caused him great difficulties in many aspects of life, including keeping
his job. Parker, Cahill, and McGaugh (2006) describe a current case of an individual
with highly detailed memory.1 She is able to remember many details from years ago in
her life but had difficulty in school and seems to perform poorly on tasks of abstract
reasoning such as processing analogies. There are many situations when we need to
rise above the details of our experience and get to their true meaning and significance.
115
1 She has written her own biography, The Woman Who Can’t Forget (Price, 2008).
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 115
116 | Representation of Knowledge
In this chapter we will address the following questions:
• What happens to our memories for gist versus detail with the passage of time? • How do we represent the gist of an experience? • Are there representations of knowledge that are not tied to specific perceptual
modalities? • What is the nature of our representation of categorical knowledge and how
does this affect the way we perceive the world?
•Knowledge and Regions of the Brain
Figure 5.1 shows some of the brain regions involved in the abstraction of
knowledge. Some prefrontal regions are associated with extracting meaningful
information from pictures and sentences. The left prefrontal region is more
involved in the processing of verbal material and the
right prefrontal region is more involved in the processing
of visual material (Gabrielli, 2001). Part of the
processing is to represent events in terms of general
categories such as bride or wedding. This categorical
information is represented in posterior regions of the
temporal, parietal, and occipital cortices. As we will see,
there is evidence that different posterior regions represent
different types of concepts.
We will review neuroscience data on the localization
of processing and information in the brain, but much
of the most striking evidence comes from behavioral
studies that examine what people remember or forget
after an event.
Prefrontal regions of the brain are associated with meaningful processing of
events, whereas posterior regions are associated with representing concepts.
•Memory for Meaningful Interpretations of Events
Memory for Verbal Information
An experiment by Wanner (1968) illustrates circumstances in which people do
and do not remember information about exact wording.Wanner asked participants
to come into the laboratory and listen to tape-recorded instructions. For
one group of participants, the warned group, the tape began this way:
The materials for this test, including the instructions, have been recorded on
tape. Listen very carefully to the instructions because you will be tested on
your ability to recall particular sentences which [sic] occur in the instructions.
Prefrontal regions
that process pictures
and sentences
Posterior regions
that represent
concepts
Brain Structures
FIGURE 5.1 Cortical regions
involved in the processing of
meaning and the representation
of concepts.
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 116
Memory for Meaningful Interpretations of Events | 117
The participants in the second group received no such warning and so had no
idea that they would be responsible for the verbatim instructions. After this
point, the instructions were the same for both groups. At a later point in the
instructions, one of four possible critical sentences was presented:
1. When you score your results, do nothing to correct your answers
but mark carefully those answers which are wrong.
2. When you score your results, do nothing to correct your answers
but carefully mark those answers which are wrong.
3. When you score your results, do nothing to your correct answers
but mark carefully those answers which are wrong.
4. When you score your results, do nothing to your correct answers
but carefully mark those answers which are wrong.
Immediately after one of these sentences was presented, all participants (warned
or not) heard the following conclusion to the instructions:
To begin the test, please turn to page 2 of the answer booklet and judge which
of the sentences printed there occurred in the instructions you just heard.
On page 2, they found the critical sentence they had just heard plus a similar
alternative. Suppose they had heard sentence 1. They might have to choose
between sentences 1 and 2 or between sentences 1 and 3. Both pairs differ only
in the ordering of two words. However, the difference between 1 and 2 does not
contribute critically to the meaning of the sentences; the difference is just stylistic.
On the other hand, sentences 1 and 3 clearly do differ in meaning. Thus, by
looking at participants’ ability to discriminate between different pairs of sentences,
Wanner was able to measure their ability to remember the meaning
versus the style of the sentence and to determine how this ability was affected
by whether or not they were warned.
The relevant data are presented in Figure 5.2. The percentage of correct
identifications of sentences heard is displayed as a function of whether participants
had been warned. The percentages are plotted separately for participants
who were asked to discriminate a meaningful difference in wording and for
those who were asked to discriminate a stylistic difference. If participants were
just guessing, they would have scored 50% correct by chance; thus, we would
not expect any values below 50%.
The implications of Wanner’s experiment are clear.
First, memory is better for changes in wording that result
in changes of meaning than for changes in wording
that result just in changes of style. The superiority of
memory for meaning indicates that people normally
extract the meaning from a linguistic message and do
not remember its exact wording. Moreover, memory for
meaning is equally good whether people are warned or
not. (The slight advantage for unwarned participants
does not approach statistical significance.) Thus, participants
retained the meaning of a message as a normal
Memory for meaning
Memory for style
Unwarned Warned
Correct (%)
50
60
70
80
90
100
FIGURE 5.2 Results from
Wanner’s experiment to
determine circumstances in
which people do and do not
remember information about
exact wording. The ability
of participants to remember
a wording difference that
affected meaning versus one
that affected only style is
plotted as a function of whether
or not the participants were
warned that they would be
tested on their ability to recall
particular sentences. (After Wanner,
1968. Adapted by permission of the author.)
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 117
118 | Representation of Knowledge
part of their comprehension process. They did not have to be cued to remember
the sentence.
The second implication of these results is that the warning did have an effect
on memory for the stylistic change. The unwarned participants remembered the
stylistic change at about the level of chance, whereas the warned participants
remembered it almost 80% of the time. This result indicates that we do not
normally retain much information about exact wording, but we can do so when
we are cued to pay attention to such information. Even with a warning, however,
memory for stylistic information is poorer than memory for meaning.
After processing a linguistic message, people usually remember just its meaning
and not its exact wording.
Memory for Visual Information
On many occasions, our memory capacity seems much greater for visual information
in a picture or a scene than for verbal information (whether that verbal
information is presented auditorially by speech or visually by text). Shepard
(1967) reported an experiment in which he had participants study a set of
magazine pictures, one picture at a time. Then they were presented with pairs
of pictures consisting of one they had studied and one they had not. The task
was to recognize which picture of each pair had been studied. This task was
contrasted with a verbal condition in which participants studied sentences and
were similarly tested on their ability to recognize those sentences when they
were presented in pairs containing one new and one studied sentence. Participants
made errors 11.8% of the time in the sentence condition but only 1.5% of
the time in the picture condition. In other words, recognition memory was
fairly high in the sentence condition but was virtually perfect in the picture
condition. There have been a number of experiments like Shepard’s, which
involved 600 pictures. Perhaps the most impressive demonstration of visual
memory is the experiment by Standing (1973), who showed that participants
had only a 17% error rate after studying 10,000 pictures.
Although people can show very good memory for pictures under some circumstances,
what they seem to be remembering is some interpretation of the
picture rather than the exact picture itself. That is, it proves useful to distinguish
between the meaning of a picture and the physical picture, just as it is
important to distinguish between the meaning of a sentence and the physical
sentence. A number of experiments point to the utility of this distinction with
respect to picture memory and to the fact that we tend to remember an interpretation
of the picture, not the physical picture.
For instance, consider an experiment by Mandler and Ritchey (1977). They
asked participants to study pictures of scenes, such as the classroom scenes in
Figure 5.3. After studying eight such pictures for 10 s each, participants were
tested for their recognition memory. They were presented with a series of pictures
and instructed to identify which pictures they had studied. The series included
the exact pictures they had studied as well as distracter pictures. A distracter such
as the one shown in Figure 5.3b was called a token distracter. It differed from
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 118
Memory for Meaningful Interpretations of Events | 119
the target only in the pattern of the teacher’s clothes, a visual detail relatively
unimportant to most interpretations of the picture. The distracter shown in
Figure 5.3c, by contrast, involves a type change—from a world map to an art
picture used by the teacher. This visual detail is relatively more important to
most interpretations of the picture because it indicates the subject being taught.
All eight pictures shown to participants contained possible token changes and
type changes. In each case, the type change involved a more important alteration
to the picture’s meaning than did the token change. There was no systematic
difference in the amount of physical change involved in a type change versus a
token change. Participants were able to recognize the original pictures 77% percent
of the time and to reject the token distracters only 60% of the time, but
they rejected the type distracters 94% of the time.
The conclusion in this study is very similar to that in the Wanner (1968)
experiment reviewed earlier. Wanner found that participants were much more
sensitive to meaning-significant changes in a sentence; Mandler and Ritchey
(1977) found that participants were more sensitive to meaning-significant
(a)
(b) (c)
FIGURE 5.3 Pictures similar to those used by Mandler and Ritchey in their experiment to
demonstrate that people distinguish between the meaning of a picture and the physical
picture itself. Participants studied the target picture (a). Later they were tested with a series
of pictures that included the target (a) along with token distracters such as (b) and type
distracters such as (c). (After Mandler & Ritchey, 1977. Adapted by permission of the publisher. © 1977 by the American
Psychological Association.)
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 119
changes in a picture. It may be that people have
better memory for the meanings of pictures than
for the meanings of sentences, but they have poor
memory for the physical details of both.
Bower, Karlin, and Dueck (1975) reported an
amusing demonstration of the fact that people’s
good memory for pictures is tied to their interpretation
of those pictures. Figure 5.4 illustrates some
of the material they used. These investigators had
participants study such drawings, called droodles,
with or without an explanation of their meaning.
Then they were given a memory test in which
they had to redraw the pictures. Participants who
had been given an explanation when studying
the pictures showed better recall (70% correctly
reconstructed) than those who were not given an explanation (51% correctly
reconstructed). Thus, memory for the drawings depended critically on participants’
ability to place a meaningful interpretation on the pictures.
When people see a picture, they tend to remember a meaningful interpretation
of it.
Retention of Detail versus Meaning
There is evidence that people initially encode many of the perceptual details of a
sentence or a picture but tend to forget this information quickly. Once the perceptual
information is forgotten, people retain memory for their interpretation of the
picture. Memory for the orientation of a picture is one of the visual details that
appear to decay rapidly, as demonstrated in an experiment by Gernsbacher
(1985). Participants were shown pictures such as the ones illustrated in Figure 5.5.
After studying one of these pictures, the participants were asked to judge which
of the pair they had seen. After 10 s, participants made their judgments with
79% accuracy, showing considerable retention of information about left–right
120 | Representation of Knowledge
(a) (b)
FIGURE 5.4 Recalling
“droodles.” (a) A midget playing
a trombone in a telephone
booth. (b) An early bird who
caught a very strong worm.
(From Bower, Karlin, & Dueck, 1975.
Reprinted by permission of the publisher.
© 1975 by Memory & Cognition.)
FIGURE 5.5 Example picture
from an experiment by
Gernsbacher, displayed in
one orientation (left) and the
reverse (right). The experiment
showed that memory for the
orientation of a picture is a
visual detail that appears to
decay rapidly. (From Gernsbacher,
1985. Original illustration from Mercer
and Mariana Meyer, One Frog Too Many.
© 1975 by Mercer and Mariana Meyer.
Reprinted by permission of the publisher,
Dial Books for Young Readers, New York.)
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 120
Memory for Meaningful Interpretations of Events | 121
orientation. After 10 min, however, their accuracy in judgment had fallen to 57%
(50% percent would reflect chance guessing). On the other hand, their memory
for what the picture was about remained high over that period of time.
When I was a graduate student, I performed an experiment (Anderson,
1974b) that made the same point in the verbal domain. Participants listened to a
story that contained various critical sentences that would be tested; for instance:
The missionary shot the painter.
Later, participants were presented with one of the following sentences and
asked whether it followed logically from the story they had heard. They were
also asked to judge which sentence they had actually heard.
1. The missionary shot the painter.
2. The painter was shot by the missionary.
3. The painter shot the missionary.
4. The missionary was shot by the painter.
The first two sentences require a positive response to the logical judgment, and
the last two require a negative response. Participants were tested either immediately
after hearing the sentence or after a delay of about 2 min. The delay had
little effect on the accuracy of their logical judgments (e.g., 1 versus 3 above)—
98% were correct immediately and 96% were correct after a delay. However,
when they were asked to judge which sentence they had heard (e.g., 1 versus 2
above), the delay had a dramatic effect. Participants were 99% correct immediately
after hearing the sentence but only 56% correct after a delay.
Memory for detail is available initially but is forgotten rapidly, whereas
memory for meaning is retained.
Implications of Good Memory for Meaning
We have seen that people have relatively good memory for meaningful interpretations
of information. So when faced with material to remember, it will
help if they can give it some meaningful interpretation. Unfortunately, many
people are unaware of this fact, and their memory performance suffers as a
consequence. I can still remember the traumatic experience I had in my first
paired-associates experiment. It happened in a sophomore class in experimental
psychology. For reasons I have long since forgotten, we had designed a
class experiment that involved learning 16 pairs, such as DAX-GIB. Our task
was to recall the second half of the pair when prompted with the first half. I was
determined to outperform the other members of my class. My personal theory
of memory at that time, which I intended to apply, was that if you try hard
and focus intensely, you can remember anything well. In the impending experimental
situation, this meant that during the learning period I should say (as
loud as was seemly) the paired associates over and over again, as fast as I could.
I believed that this method would burn the paired associates into my mind
forever. To my chagrin, I wound up with the worst score in the class.
Anderson7e_Chapter_05.qxd 8/20/09 9:43 AM Page 121
My theory of “loud and fast” was directly opposed to the true means of
improving memory. I was trying to memorize a meaningless verbal pair. But
the material discussed in this chapter so far suggests that we have the best
memory for meaningful information. I should have been trying to convert my
memory task into something more meaningful. For instance, DAX is like dad
and GIB is the first part of gibberish. So I might have created an image of my
father speaking some gibberish to me. This would have been a simple mnemonic
(memory-assisting) technique and would have worked quite well as a means of
associating the two elements.
We do not often need to learn pairs of nonsense syllables outside the laboratory.
In many situations, however,we do have to associate various combinations of
terms that do not have much inherent meaning.We have to remember shopping
lists, names for faces, telephone numbers, rote facts in a college class, vocabulary
items in a foreign language, and so on. In all cases, we can improve memory if we
associate the items to be remembered with a meaningful interpretation.
122 | Representation of Knowledge
effectiveness of this technique (for a review, read Kroll &
DeGroot, 2005). The research shows that, like many
things, one needs to take a nuanced approach in evaluating
the effectiveness of the keyword technique. There is
no doubt that it results in more rapid
vocabulary learning in many situations,
but there are potential costs. One might
imagine that having to go through the
intermediate keyword slows down
speed of translation, and the keyword
method has been shown to result in
slower retrieval times compared to
retrieval of items that are directly
associated without an intermediate.
Moreover, one might wonder about the
implications of having to go through
the intermediate for long-term retention,
and again it has been shown to
result in poorer long-term retention.
Finally, there is evidence that suggests
that although the method may help in passing the immediate
vocabulary test in a class and hurt in a delayed test
that we have not studied for, its ultimate impact on
achieving real language mastery is minimal. Chapter 12 will
discuss issues involved in foreign language mastery.
Implications
Mnemonic techniques for remembering vocabulary items
One domain where we seem to have to learn arbitrary
associations is foreign language vocabulary. For instance,
consider trying to learn that the Italian formaggio
(pronounced “for modge jo”) means cheese. There is a
memorization technique, called the keyword
method, for learning vocabulary
items, which some students are taught
and others discover on their own. The
first step is to convert the foreign word
to some sound-alike in one’s native
language. For example, one might convert
formaggio into “for much dough.”
The second step is to create a meaningful
connection between the sound-alike
and the meaning. For example, we might
imagine expensive cheese being sold for
much money or “for much dough.” Or
consider the Italian carciofi (pronounced
“car-choh-fee”), which means artichoke.
We might transform “car-choh-fee” into
“car trophy” and imagine a winning car at an auto show
with a trophy shaped like an artichoke. The intermediate
(e.g., for much dough or car trophy) is called the keyword,
although in both of these examples they are really
key phrases. There has been extensive research on the
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 122
Propositional Representations | 123
It is easier to commit arbitrary associations to memory if they are converted
into something more meaningful.
•Propositional Representations
We have shown that in many situations people do not remember the exact
physical details of what they have seen or heard but rather the “meaning” of
what they have encountered. In an attempt to become more precise about what
is meant by “meaning,” cognitive psychologists developed what is called a
propositional representation. The concept of a proposition, borrowed from
logic and linguistics, is central to such analyses. A proposition is the smallest
unit of knowledge that can stand as a separate assertion, that is, the smallest
unit one can meaningfully judge as true or false. Propositional analysis applies
most clearly to linguistic information, and I will develop the topic here in terms
of such information.
Consider the following sentence:
Lincoln, who was president of the United States during a bitter war, freed the slaves.
The information conveyed in this sentence can be communicated by the following
simpler sentences:
A. Lincoln was president of the United States during a war.
B. The war was bitter.
C. Lincoln freed the slaves.
If any of these simple sentences were false, the complex sentence also would
be false. These sentences correspond closely to the propositions that underlie
the meaning of the complex sentence. Each simple sentence expresses a primitive
unit of meaning. One condition that our meaning representations must satisfy is
that each separate unit composing them must correspond to a unit of meaning.
However, the theory of propositional representation does not claim that a
person remembers the simple sentences such as the three just presented. Past
research indicates that people do not remember the exact wording of such
simple sentences any more than they remember the exact wording of the
original complex sentence. For instance, in another study I did as a graduate
student (Anderson, 1972), I showed that participants demonstrated poor ability
to remember whether they had heard sentence C or the sentence:
The slaves were freed by Lincoln.
Thus, it seems that we represent information in memory in a way that preserves
the meaning of the primitive assertions but does not preserve any information
about specific wording. A number of propositional notations represent information
in this abstract way. One, used by Kintsch (1974), represents each
proposition as a list containing a relation followed by an ordered list of
arguments. The relations organize the arguments and typically correspond
to the verbs (in this case, free), adjectives (bitter), and other relational terms
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 123
(president of ). The arguments refer to particular times, places, people, or objects,
and typically correspond to the nouns (Lincoln, war, slaves). The relations
assert connections among the entities these nouns refer to. Kintsch represents
each proposition by a parenthesized list consisting of a relation plus arguments.
As an example, sentences A through C would be represented by these lists:
a. (president-of: Lincoln, United States, war)
b. (bitter: war)
c. (free: Lincoln, slaves)
Note that each relation takes a different number of arguments: president of takes
three, free takes two, and bitter takes one. Whether a person heard the original
complex sentence or heard
The slaves were freed by Lincoln, the president during a bitter war,
the meaning of the message would be represented by lists a through c.
Bransford and Franks (1971) provided an interesting demonstration of the
psychological reality of propositional units. In this experiment, participants
studied 12 sentences, including the following:
The ants ate the sweet jelly, which was on the table.
The rock rolled down the mountain and crushed the tiny hut.
The ants in the kitchen ate the jelly.
The rock rolled down the mountain and crushed the hut beside the woods.
The ants in the kitchen ate the jelly, which was on the table.
The tiny hut was beside the woods.
The jelly was sweet.
All these sentences are composed from two sets of four propositions. One set of
four propositions can be represented as
1. (eat: ants, jelly, past)
2. (sweet: jelly)
3. (on: jelly, table, past)
4. (in: ants, kitchen, past)
The other set of four propositions can be represented as
1. (roll down: rock, mountain, past)
2. (crush: rock, hut, past)
3. (beside: hut, woods, past)
4. (tiny: hut)
Bransford and Franks looked at participants’ recognition memory for the following
three kinds of sentences:
1. Old: The ants in the kitchen ate the jelly.
2. New: The ants ate the sweet jelly.
3. Noncase: The ants ate the jelly beside the woods.
124 | Representation of Knowledge
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 124
Propositional Representations | 125
The first kind of sentence was actually studied, the second was not but is a combination
of propositions that were studied, and the third consists of words that
were studied but cannot be composed from the propositions studied. Bransford
and Franks found that participants had almost no ability to discriminate between
the first two kinds of sentences and were equally likely to say that they
had actually heard either. On the other hand, participants were quite confident
that they had not heard the third, noncase, sentence.
The experiment shows that although people remember the propositions they
encounter, they are quite insensitive to the actual combination of propositions.
Indeed, the participants in this experiment were most likely to say that they
heard a sentence consisting of all four propositions, such as
The ants in the kitchen ate the sweet jelly, which was on the table,
even though they had not in fact studied this sentence.
According to propositional analyses people remember a complex sentence as a
set of abstract meaning units that represent the simple assertions in the sentence.
Propositional Networks
In the cognitive psychology literature, one sometimes finds propositions represented
in a network form. Figure 5.6 illustrates the structure of a propositional
network that encodes the sentence, “Lincoln, who was president of
(a) (b) (c)
War
United States
President-of
Time Agent
Lincoln
Object
Relation
Lincoln
Freed
Object
Slaves
Agent
Relation
War
Bitter
Subject
Relation
(d)
War
United States
President-of
Time Agent
Lincoln
Object
Relation
Freed
Object
Slaves
Agent
Relation
Bitter
Subject
Relation
FIGURE 5.6 Network
representations for the
proposition underlying the
statement: “Lincoln, who was
president of the United States
during a bitter war, freed
the slaves.”
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 125
the United States during a bitter war, freed the slaves.” In this propositional
network, each proposition is represented by an ellipse, which is connected by
labeled arrows to its relation and arguments. The propositions, the relations,
and the arguments are called the nodes of the network, and the arrows are
called the links because they connect the nodes. For instance, the ellipse in
Figure 5.6a represents proposition (a) from the earlier Kintsch analysis. This
ellipse is connected to the relation president-of by a link labeled relation (to
indicate that it is pointing to the relation node), to Lincoln by an agent link,
to United States by an object link, and to war by a time link. The three network
structures shown in Figures 5.6a, 5.6b, and 5.6c represent the individual
propositions (a) through (c) from the earlier Kintsch analysis. Note that these
three networks contain the same nodes, for example, Figures 5.6a and 5.6b
both contain war. This overlap indicates that these networks are really interconnected
parts of a larger network, which is illustrated in Figure 5.6d. This
last network represents all the meaningful information in the original complex
sentence on page 147.
The spatial location of elements in a network is irrelevant to its interpretation.
A network can be thought of as a tangle of marbles connected by strings.
The marbles represent the nodes, the strings the links between nodes. The network
represented on a 2-D page is that tangle of marbles laid out in a certain
way. We try to lay out the network in a way that facilitates an understanding
of it, but any layout is possible. All that matters is which elements are connected
to which others, not where the components lie.
A number of experiments suggest that it is helpful to think of the nodes
in such networks as ideas and to think of the links between the nodes as
associations between the ideas. Consider an experiment by Weisberg (1969)
that used a constrained association task. In this experiment, participants studied
and committed to memory such sentences as “Children who are slow eat bread
that is cold.” The propositional network representation of this sentence is illustrated
in Figure 5.7. After learning a sentence, participants were administered
free-association tasks in which they were given a word from the sentence and
asked to respond with the first word from the sentence that came to mind.
Participants cued with slow almost always free-associated children and almost
126 | Representation of Knowledge
Relative Subject
Slow Children Bread Cold
Agent Object
Time Relation
Past Eat
Subject Relation
FIGURE 5.7 A propositional network representation of the sentence: “Children who are slow
eat bread that is cold.” Weisberg (1969) used such sentences in an experiment to show that
the proximity of words in a propositional network has more effect on memory than their
physical proximity in the sentence.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 126
Propositional Representations | 127
never bread, although bread is closer to slow in the sentence than to children.
However, the illustration shows that slow and children are nearer each other in
the network (two links) than slow and bread (four links). Similarly, participants
cued with bread almost always recalled cold rather than slow, although in
the sentence bread and slow are closer than bread and cold. Again, bread and
cold are closer to each other in the network (three links) than are bread and slow
(five links). (A similar point was made in an experiment by R. A. Ratcliff &
McKoon, 1978.)
Propositional information can be represented in networks that display how
concepts relate.
Amodal versus Perceptual Symbol Systems
The propositional representations that we have just considered are examples of
what Barsalou (1999) called an amodal symbol system. By this he meant that
the elements within the system are inherently nonperceptual. The original stimulus
might be a picture or a sentence, but the representation is abstracted away
from the verbal or visual modality. Given this abstraction, one would predict
that participants in experiments would be unable to remember the exact words
they heard or the exact picture they saw.
As an alternative to such theories, Barsalou proposed a hypothesis called
the perceptual symbol system, which suggests that all information is represented
in terms that are specific to a particular perceptual modality (visual,
auditory, etc.) and basically perceptual. The perceptual symbol hypothesis is
an extension of Paivio’s (1971, 1986) dual-code theory that claimed that,
rather than abstract propositional representations, we represent information
in combined verbal and visual codes. Paivio suggested that when we hear a
sentence, we develop an image of what it describes. If we later remember the
visual image and not the sentence, we will remember what the sentence was
about, but not its exact words. Analogously, when we see a picture, we might
describe to ourselves the significant features of that picture. If we later remember
our description and not the picture, we will not remember details we did
not think important to describe (such as the clothes the teacher was wearing in
Figure 5.3).
The dual-code position does not predict that memory for the wording of a
sentence is necessarily poor. The relative memory for the wording versus memory
for the meaning depends on the relative attention that people give to the
verbal versus the visual representation. There are a number of experiments
showing that when participants pay attention to wording, they show better
memory. For instance, Holmes,Waters, and Rajaram (1998), in a replication of
the Bransford and Franks (1971) study that we just reviewed, asked participants
to count the number of letters in the last word of each sentence. This manipulation,
which increased their attention to the wording of the sentence, resulted
in an increased ability to discriminate sentences they had studied from sentences
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 127
with similar meanings that they had not—although participants still showed
considerable confusion among similar-meaning sentences.
But how can an abstract concept such as honesty be represented in a purely
perceptual cognitive system? One can be very creative in combining perceptual
representations. Consider a pair of sentences from an old unpublished study of
mine.2We had participants study one of the following two sentences:
1. The lieutenant wrote his signature on the check.
2. The lieutenant forged a signature on the check.
Later, we asked them to recognize which sentence they had studied. They could
make such discriminations more successfully than they could distinguish
between pairs such as
1. The lieutenant enraged his superior in the barracks.
2. The lieutenant infuriated a superior in the barracks.
In the first pair of sentences, there is a big difference in meaning; in the second
pair, little difference. However, the difference in wording between the sentences
in the two pairs is equivalent. When I did the study, I thought it showed that
people could remember meaning distinctions that did not have perceptual
differences—the distinction between signing a signature and forging is not in
what the person does but in his or her intentions and the relationship between
those intentions and unseen social contracts. Barsalou (personal communication,
March 12, 2003) suggested that we represent the distinction between the two sentences
by reenacting the history behind each sentence. So even if the actual act
of writing and forging might be the same, the history of what a person said and
did in getting to that point might be different. Barsalou also considers the internal
state of the individual to be relevant. Thus, part of the perceptual features
involved in forging might include the sensations of tension that one has when
one is in a difficult situation.3
Barsalou, Simmons, Barbey, and Wilson (2003) cited evidence that when
people understand a sentence, they actually come up with a perceptual interpretation
of that sentence. For instance, in one study by Stanfield and Zwaan
(2001), participants read a sentence about a nail being pounded either into the
wall or the floor. Then they viewed a picture of a nail oriented either horizontally
or vertically and were asked to affirm whether the object in the picture was
mentioned in the sentence that they just read. If they had read a sentence about
a nail being pounded into the wall, they recognized a horizontally oriented nail
more quickly. When they had read a sentence about a nail being pounded into
the floor, they recognized a vertically oriented nail more quickly. In other words,
they responded faster when the orientation implied by the sentence matched the
orientation of the picture. Thus, their interpretation of the sentence seemed to
128 | Representation of Knowledge
2 It was not published because at the time (1970s) it was considered too obvious a result given studies like
those described earlier in this chapter.
3 Perhaps it is obvious that I do not agree with Barsalou’s perspective. However, it is hard to imagine what
he might consider disconfirming data, because his approach is so flexible.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 128
contain this perceptual detail. As further evidence of the perceptual representation
of meaning, Barsalou et al. cited neuroscience studies showing that concepts
are represented in brain areas similar to those that process perceptions.
An alternative to amodal representations of meaning is the view that meaning
is represented as a combination of images in different perceptual modalities.
•Embodied Cognition
The perceptual symbol hypothesis of Barsalou is an instance of the growing
emphasis in psychology on understanding the contribution of the environment
and our bodies to shaping our cognition. As Thelen (2000) describes the
viewpoint:
To say that cognition is embodied means that it arises from bodily interactions
with the world and is continually meshed with them. From this point of view,
therefore, cognition depends on the kinds of experiences that come from having
a body with particular perceptual and motor capabilities that are inseparably
linked and that together form the matrix within which reasoning, memory,
emotion, language and all other aspects of mental life are embedded. (p. 5)
The embodied cognition perspective emphasizes the contribution of motor
action and how it connects us to the environment. For instance, Glenberg (2007)
argues that our understanding of language often depends on covertly acting out
what the language describes. He points to an fMRI study by Hauk, Johnsrude, &
Pulvermiller (2004), who recorded brain activation while people listened to verbs
that involved the face, arm, or leg actions (e.g., to lick, pick, or kick). They looked
for activity along the motor cortex in different regions associated with the face,
arm, and leg (see Figure 1.10). Figure 5.8 shows the differential activity in these
Embodied Cognition | 129
Region: Face
MR signal change (arbitrary units)
Arm Leg
0.02
0.04
0.06
0.08
0.1
Leg words
Face words
Arm words
FIGURE 5.8 Brain activation in different model regions as participants listen to different types
of verbs.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 129
different brain regions. As participants listened to each word, there was greater
activation in the part of the motor cortex that would produce that action.
A theory of how meaning is represented in the human mind must explain
how different perceptual and motor modalities connect with one another. For
instance, part of understanding a word like kick is our ability to relate it to a
picture of a person kicking a ball so that we can describe that picture. As another
example, part of our understanding of someone performing an action is
our ability to relate to our own motor system so that we can mimic the action.
Interestingly, mirror neurons have been found in the motor cortex of monkeys,
which are active when the monkeys perform an action like ripping a paper, or
see the experimenter rip a paper or hear the experimenter rip the paper without
seeing the action (Rizzolatti & Craighero, 2004). Although one cannot typically
do single-cell recordings with humans, brain-imaging studies have found
increased activity in the motor region when people observe actions, particularly
with the intention to mimic the action (Iacoboni et al., 1999).
Figure 5.9 illustrates two conceptions of how mappings might take place
between different representations. One possibility is illustrated in the multimodal
hypothesis, which holds that we have various representations tied to
different perceptual and motor systems and that we have means of directly
converting one representation to another. For instance, the double-headed
arrow going from the visual to the motor would be a system for converting a
visual representation into a motor representation and a system for converting
the representations in the opposite direction. The alternative amodal hypothesis
is that there is an intermediate abstract system, perhaps the propositional representation
that we described earlier, and that we have systems for converting
back and forth between the perceptual and motor representations and this
abstract representation. So to convert a picture into an action, one first converts
the visual representation into an abstract representation of its significance
and then converts that representation into a motor representation. These two
approaches offer alternative explanations for the research we reviewed earlier
that indicated people remember the meaning of what they experience, but not
130 | Representation of Knowledge
Multimodal Hypothesis Amodal Hypothesis
Visual
Other
Verbal
Motor
Visual
Other
Verbal
Motor
Meaning
FIGURE 5.9 Representations of two hypotheses about how information is related between
different perceptual and motor modalities. The multimodal hypothesis holds that there are
mechanisms for translating between each modality. The amodal hypothesis holds that
each modality can be translated back and forth to a central meaning representation.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 130
Conceptual Knowledge | 131
the details. The amodal hypothesis holds that this information is retained in the
central meaning system. The multimodal hypothesis holds that the person has
converted the information from the modality of the presentation to some other
modality.
The embodied cognition perspective emphasizes that meaning is represented
in the perceptual and motor systems that we use to interact with the world.
•Conceptual Knowledge
Consider the picture in Figure 5.3a.When we look at this picture, we do not see
it as just a collection of specific objects. Rather, we see it as a picture of a teacher
instructing a student on geography. That is, we see the word in terms of categories
like teacher, student, instruction, and geography. As we saw, people tend to
remember this categorical information and not the specific details. For instance,
the participants in the Mandler and Ritchey (1977) experiment forgot what the
teacher wore but remembered the subject she taught.
You cannot help but experience the world in terms of the categories you
know. For instance, if you were licked by a four-legged furry object that weighed
about 50 pounds and had a wagging tail, you would perceive yourself as being
licked by a dog.What does your cognitive system gain by categorizing the object
as a dog? Basically, it gains the ability to predict. Thus, you can have expectations
about what sounds this creature might make and what would happen if you
threw a ball (the dog might chase it and stop licking you). Because of this ability
to predict, categories give us great economy in representation and communication.
For instance, if you tell someone, “I was licked by a dog,” your listener can
predict the number of legs on the creature, its approximate size, and so on.
The effects of such categorical perceptions are not always positive—for instance,
they can lead to stereotyping. In one study, Dunning and Sherman
(1997) had participants study sentences like
Elizabeth was not very surprised upon receiving her math SAT score.
or
Bob was not very surprised upon receiving his math SAT score.
Participants who had heard the first sentences were more likely to falsely believe
they had heard “Elizabeth was not very surprised upon receiving her low math
SAT score,” whereas if they had heard the second sentence, they were more
likely to believe they had heard “Bob was not very surprised upon receiving his
high math SAT score.” Categorizing Elizabeth as a woman, the participants
brought the stereotype of women as poor at math to their interpretation of the
first sentence. Categorizing Bob as male, they brought the opposite stereotype
to their interpretation of the second sentence. This was even true among participants
(both male and female) who were rated as not being sexist in their
attitudes. They could not help but be influenced by their implicit stereotypes.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 131
Research on categorization has focused both on how we form these categories
in the first place and on how we use them to interpret experiences. It has
also been concerned with notations for representing this categorical knowledge.
In this section, we will consider a number of proposed notations for representing
conceptual knowledge. We will start by describing two early theories. One
concerns semantic networks, which are similar to the propositional networks
we just considered. The other is about what are called schemas. Both theories
have been closely related to certain empirical phenomena that seem central to
conceptual structure.
The categorical organization of our knowledge strongly influences the way
we encode and remember our experiences.
Semantic Networks
Network representations have been used to encode conceptual knowledge as
well as propositional knowledge. Quillian (1966) proposed that people store
information about various categories—such as canaries, robins, fish, and so
on—in a network structure like that shown in Figure 5.10. In this illustration, we
represent a hierarchy of categorical facts, such as that a canary is a bird and a bird
is an animal, by linking nodes for the two categories with isa links. Properties
that are true of the categories are associated with them. Properties that are true
of higher-level categories are also true of lower level categories. Thus, because
animals breathe, it follows that birds and canaries breathe. Figure 5.10 can also
132 | Representation of Knowledge
Level 1
Level 2
Level 3 Canary
Can sing
Is yellow
Bird
Has wings
Can fly
Has feathers
Animal
Has skin
Can move around
Eats
Breathes
Ostrich
Has long
thin legs
Is tall
Can’t fly
Shark
Can bite
Is dangerous
Fish
Has fins
Can swim
Has gills
Salmon
Is pink
Is edible
Swims
upstream
to lay eggs
FIGURE 5.10 A hypothetical memory structure for a three-level hierarchy using the example
canary. Quillian (1966) proposed that people store information about various categories in a
network structure. This illustration represents a hierarchy of categorical facts, such as that a
canary is a bird and a bird is an animal. Properties that are true of each category are associated
with that category. Properties that are true of higher level categories are also true of lower level
categories. (After Collins & Quillian, 1969. Adapted by permission of the publisher. © 1969 by Academic Press.)
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 132
Conceptual Knowledge | 133
represent information about exceptions. For instance, even though most birds
fly, the illustration does represent that ostriches cannot fly.
Collins and Quillian (1969) did an experiment to test the psychological reality
of such networks by having participants judge the truth of assertions about
concepts, such as
1. Canaries can sing.
2. Canaries have feathers.
3. Canaries have skin.
Participants were shown these along with false assertions such as “Apples have
feathers.” They were asked to indicate whether a statement was true or false by
pressing one of two buttons. The time from presentation of the statement to
the button press was measured.
Consider how participants would answer such questions if Figure 5.10 represented
their knowledge of such categories. The information to confirm sentence
1 is directly stored with canary. The information for sentence 2, however,
is not directly stored at the canary node. Instead, the have feathers property is
stored with bird, and sentence 2 can be inferred from the directly stored facts
that A canary is a bird and Birds have feathers. Again, sentence 3 is not directly
stored with canary; rather, the has skin predicate is stored with animal. Thus,
sentence 3 can be inferred from the facts that a canary is a bird and a bird is an
animal and animals have skin. So, all the information required to verify sentence
1 is stored with canary; for sentence 2, participants would need to traverse
one link, from canary to bird, to retrieve the requisite information; for sentence 3,
they would have to traverse two links, from canary to animal.
If our categorical knowledge were structured like Figure 5.10, we would
expect sentence 1 to be verified more quickly than sentence 2, which would be
verified more quickly than sentence 3. This is just what Collins and Quillian
found. Participants required 1310 ms to judge statements like sentence 1, 1380
ms to judge statements like sentence 2, and 1470 ms to judge statements like
sentence 3. Subsequent research on the retrieval of information from memory
has somewhat complicated the conclusions drawn from the initial Collins and
Quillian experiment. How often facts are experienced has been observed to
have strong effects on retrieval time (e.g., C. Conrad, 1972). Some facts, such as
Apples are eaten—for which the predicate could be stored with an intermediate
concept such as food, but that are experienced quite often—are verified as fast
as or faster than facts such as Apples have dark seeds, which must be stored more
directly with the apple concept. It seems that if a fact about a concept is encountered
frequently, it will be stored with that concept, even if it could also
be inferred from a more general concept. The following statements about the
organization of facts in semantic memory and their retrieval times seem to be
valid conclusions from the research:
1. If a fact about a concept is encountered frequently, it will be stored
with that concept even if it could be inferred from a higher order
concept.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 133
2. The more frequently a fact about a concept is encountered, the more
strongly that fact will be associated with the concept. The more strongly
facts are associated with concepts, the more rapidly they are verified.
3. Inferring facts that are not directly stored with a concept takes a
relatively long time.
Thus, both the strength of the connections between facts and concepts (determined
by frequency of experience) and the distance between them in the
semantic network have effects on retrieval time.
When a property is not stored directly with a concept, people can retrieve it
from a higher order concept.
Schemas
Consider our knowledge of what a house is like. We know many things about
houses, such as
• Houses are a type of building. • Houses have rooms. • Houses can be built of wood, brick, or stone. • Houses serve as human dwellings. • Houses tend to have rectilinear and triangular shapes. • Houses are usually larger than 100 square feet and smaller than
10,000 square feet.
The importance of a category is that it stores predictable information about
specific instances of that category. So when someone mentions a house, for
example, we have a rough idea of the size of the object being referred to.
Semantic networks, which just store properties with concepts, cannot capture
the nature of our general knowledge about a house, such as its typical size
or shape. Researchers in cognitive science (e.g., Rumelhart & Ortony, 1976)
proposed a particular way of representing such knowledge that seemed more
useful than the semantic network representation. Their representational structure
is called a schema. The concept of a schema was first articulated in AI and
computer science. Readers who have experience with modern programming
languages should recognize its similarity to various types of data structures.
The question for the psychologist is what aspects of the schema notion are
appropriate for understanding how people reason about concepts. I will describe
some of the properties associated with schemas and then discuss the psychological
research bearing on these properties.
Schemas represent categorical knowledge according to a slot structure, in
which slots specify values of various attributes that members of a category possess.
So we have the following partial schema representation of a house:
House • Isa: building • Parts: rooms
134 | Representation of Knowledge
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 134
Conceptual Knowledge | 135
• Materials: wood, brick, stone • Function: human dwelling • Shape: rectilinear, triangular • Size: 100–10,000 square feet
In this representation, such terms as materials and shape are the attributes or
slots, and such terms as wood, brick, and rectilinear are the values. Each pair of a
slot and a value specifies a typical feature. The fact that houses are usually built
of materials such as wood and brick does not exclude such possibilities as cardboard.
Thus, the values listed above are called default values. For instance, the
fact that we represent that birds can fly as part of our schema for birds does not
prevent us from seeing ostriches as birds.We simply overwrite this default value
in our representation of an ostrich.
A special slot in each schema is its isa slot, which is like the isa link in a
semantic network and points to the superset. Basically, unless contradicted, a
concept inherits the features of its superset. Thus, with the schema for building,
the superset of house, we would store such features as that it has a roof and
walls and that it is found on the ground. This information is not represented
in the schema for house because it can be inferred from building. As illustrated
in Figure 5.10, these isa links can create a structure called a generalization
hierarchy.
Schemas have another type of structure, called a part hierarchy. Parts of
houses, such as walls and rooms, have their own schema definitions. Stored
with schemas for walls and rooms would be the information that they have windows
and ceilings as parts. Thus, using the part hierarchy, we would be able to
infer that houses have windows and ceilings.
Schemas are abstractions from specific instances that can be used to make
inferences about instances of the concepts they represent. If we know something
is a house, we can use the schema to infer that it is probably made of
wood or brick and that it has walls, windows, and ceilings. The inferential
processes for schemas must also be able to deal with exceptions: We can
still understand what a house without a roof is. Finally, it is necessary to
understand the constraints between the slots of a schema. If we hear of a
house that is underground, for example, we can infer that it will not have
windows.
Schemas represent concepts in terms of supersets, parts, and other attributevalue
pairs.
Psychological Reality of Schemas
One property of schemas is that they have default values for certain slots or
attributes. This property provides schemas with a useful inferential mechanism.
If you recognize an object as being a member of a certain category, you can
infer—unless explicitly contradicted—that it has the default values associated
with that concept’s schema. Brewer and Treyens (1981) provided an interesting
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 135
demonstration of the effects of schemas on memory inferences.
Thirty participants were brought individually to the room
shown in Figure 5.11. Each was told that this room was the office
of the experimenter and was asked to wait there while the experimenter
went to the laboratory to see whether the previous participant
had finished. After 35 s, the experimenter returned and
took the waiting participant to a nearby seminar room. Here, the
participant was asked to write down everything he or she could
remember about the experimental room. What would you be
able to recall?
Brewer and Treyens predicted that their participants’ recall
would be strongly influenced by their schema of what an office
contains. Participants would recall very well items that are part
of that schema; they would recall much less well office items that
are not part of the schema; and they would falsely recall items
that are part of the schema but not in this office. Brewer and Treyens found just
this pattern of results. For instance, 29 of the 30 participants recalled that the
office had a chair, a desk, and walls. Only 8 participants, however, recalled that
it had a bulletin board or a skull. On the other hand, 9 participants recalled
that it had books, which it did not. Thus, we see that a person’s memory for the
properties of a location is strongly influenced by that person’s default assumptions
about what is typically found in the location. A schema is a way of encoding
those default assumptions.
People will infer that an object has the default values for its category, unless
they explicitly notice otherwise.
Degree of Category Membership
One of the important features of schemas is that they allow variation in the
objects that might fit a particular schema. There are constraints on what typically
occupies the various slots of a schema, but few absolute prohibitions. Thus,
if schemas encode our knowledge about various object categories, we ought to
see a shading from less typical to more typical members of the category as the
features of the members better satisfy the schema constraints. There is now considerable
evidence that natural categories such as birds have the kind of structure
that would be expected of a schema.
Rosch did early research documenting such variations in category membership.
In one experiment (Rosch, 1973), she instructed participants to rate the
typicality of various members of a category on a 1 to 7 scale, where 1 meant
very typical and 7 meant very atypical. Participants consistently rated some
members as more typical than others. In the bird category, robin got an average
rating of 1.1, and chicken a rating of 3.8. In reference to sports, football was
thought to be very typical (1.2), whereas weight lifting was not (4.7). Murder
was rated a very typical crime (1.0), whereas vagrancy was not (5.3). Carrot was
a very typical vegetable (1.1); parsley was not (3.8).
136 | Representation of Knowledge
FIGURE 5.11 The “office room”
used in the experiment of
Brewer and Treyens to demonstrate
the effects of schemas
on memory inferences. As they
predicted, their participants’
recall was strongly influenced by
their schema of what an office
contains. (From Brewer & Treyens, 1981.
Reprinted by permission of the publisher.
© 1981 by Cognitive Psychology.)
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 136
Conceptual Knowledge | 137
Rosch (1975) also asked participants to identify the category of pictured
objects. People are faster to judge a picture as an instance of a category when it
presents a typical member of the category. For instance, apples are seen as fruits
more rapidly than are watermelons, and robins are seen as birds more rapidly
than are chickens. Thus, typical members of a category appear to have an
advantage in perceptual recognition as well.
Rosch (1977) demonstrated another way in which some members of a category
are more typical. She had participants compose sentences for category
names. For bird, participants generated sentences such as
I heard a bird twittering outside my window.
Three birds sat on the branch of a tree.
A bird flew down and began eating.
Rosch replaced the category name in these sentences with a typical member
(robin), a less typical member (eagle), or a peripheral member (chicken) and
asked participants to rate the sensibleness of the resulting sentences. Sentences
involving typical members got high ratings, sentences with less typical members
got lower ratings, and sentences with peripheral members got the lowest
ratings. So the evidence shows that when people think of a category member,
they generally think of typical instances of that category.
Failing to have a default or typical value does not disqualify an object from
being a member of the category, however. People should have great difficulty
and should be inconsistent in judging whether items at the periphery of a category
are actually members of that category. McCloskey and
Glucksberg (1978) looked at people’s judgments about what
were or were not members of various categories. They found that
although participants did agree on some items, they disagreed on
many. For instance, whereas all 30 participants agreed that cancer
was a disease and happiness was not, 16 thought stroke was a disease
and 14 did not. Again, all 30 participants agreed that apple
was a fruit and chicken was not, but 16 thought pumpkin was a
fruit and 14 disagreed. Once again, all participants agreed that a
fly was an insect and a dog was not, but 13 participants thought
a leech was and 17 disagreed. Thus, it appears that people do not
always agree among themselves.McCloskey and Glucksberg tested
the same participants a month later and found that many had
changed their minds about the disputed items. For instance, 11
out of 30 reversed themselves on stroke, 8 reversed themselves
on pumpkin, and 3 reversed themselves on leech. Thus, disagreement
about category boundaries does not occur just among
participants—people are very uncertain within themselves exactly
where the boundaries of a category should be drawn.
Figure 5.12 illustrates a set of materials used by Labov (1973).
He was interested in studying which items participants would call
cups and which they would not.Which do you consider to be cups
17
19
16
18
12 15
11 14
10 13
2 3 4
FIGURE 5.12 The various
cuplike objects used in Labov’s
experiment that studied the
boundaries of the cup category.
(After Labov, 1973, in Bailey & Shuy, 1973.
Adapted by permission of the author.
© 1973 by Georgetown University Press.)
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 137
and which do you consider bowls? The interesting
point is that these concepts do not appear to have
clear-cut boundaries. In one experiment, Labov
used the series of items 1 through 4 shown in Figure
5.12 and a fifth item, not shown. These items
reflect an increasing ratio of width of the cup to
depth. For the first item, that ratio is 1, whereas for
item 4 it is 1.9. The ratio for the item not shown
was 2.5. Figure 5.13 shows the percentage of participants
who called each of the five objects a cup
and the percentage who called it a bowl. The solid
lines indicate the classifications when participants
were simply presented with pictures of the objects
(the neutral context). As can be seen, the percentages
of cup responses gradually decreased with increasing
width, but there is no clear-cut point where participants stopped using
cup. At the extreme 2.5-width ratio, about 25% percent of the participants still
gave the cup response, whereas another 25% gave bowl. (The remaining 50%
gave other responses.) The dashed lines reflect classifications when participants
were asked to imagine the object filled with mashed potatoes and placed on a
table. In this context, fewer cup responses and more bowl responses were given,
but the data show the same gradual shift from cup to bowl. Thus, it appears that
people’s classification behavior varies continuously not only with the properties
of an object but also with the context in which the object is imagined or presented.
These influences of perceptual features and context on categorization
judgments are very much like the similar influences of these features on perceptual
pattern recognition (see Chapter 2).
Different instances are judged to be members of a category to different
degrees, with the more typical members of a category having an advantage
in processing.
Event Concepts
It is not only objects that have a conceptual structure.We also have concepts of
various kinds of events, such as going to a movie. Schemas have been proposed
as ways of representing such categories. We can encode our knowledge about
stereotypic events according to their parts—for instance, going to a movie involves
going to the theater, buying the ticket, buying refreshments, seeing the
movie, and returning from the theater. Schank and Abelson (1977) proposed
versions of event schemas that they called scripts. They pointed out that many
circumstances involve stereotypic sequences of actions. For instance, Table 5.1
shows their hunch as to what the stereotypic aspects of dining at a restaurant
might be and represents the components of a script for such an occasion.
Bower, Black, and Turner (1979) reported a series of experiments in which
the psychological reality of the script notion was tested. They asked participants
to name what they considered the 20 most important events in an episode,
138 | Representation of Knowledge
1.0 1.2
00
25
1.5 1.9 2.5
50
75
100
Cup
Bowl
Bowl
Cup
Neutral context
Food context
Relative width of cup
Response (%)
FIGURE 5.13 Results from
Labov’s experiment demonstrating
that the cup category does
not appear to have clear-cut
boundaries. The percentage of
participants who used the term
cup versus the term bowl to
describe the objects shown in
Figure 5.10 are plotted as a
function of the ratio of cup
width to cup depth. The solid
lines reflect the neutral-context
condition, the dashed lines the
food-context condition. (After
Labov, 1973, in Bailey & Shuy, 1973.
Adapted by permission of the author.
© 1973 by Georgetown University Press.)
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 138
Conceptual Knowledge | 139
such as going to a restaurant.With 32 participants, they failed to get complete
agreement on what these events were. No particular action was listed as part of
the episode by all participants, although considerable consensus was reported.
Table 5.2 lists the events named. The items in roman type were listed by at least
25% of the participants; the italicized items were named by at least 48%; and
the boldfaced items were given by at least 73%. Using 73% as a criterion, we
find that the stereotypic sequence was sit down, look at menu, order meal, eat
food, pay bill, and leave.
Bower et al. (1979) went on to show a number of the effects that such action
scripts have on memory for stories. They had participants study stories that included
some but not all of the typical events from a script. Participants were
then asked to recall the stories (in one experiment) or to recognize whether
TABLE 5.1
The Schema for Going
to a Restaurant
Scene I: Entering
Customer enters restaurant
Customer looks for table
Customer decides where to sit
Customer goes to table
Customer sits down
Scene 2: Ordering
Customer picks up menu
Customer looks at menu
Customer decides on food
Customer signals waitress
Waitress comes to table
Customer orders food
Waitress goes to cook
Waitress gives food order to cook
Cook prepares food
Scene 3: Eating
Cook gives food to waitress
Waitress brings food to customer
Customer eats food
Scene 4: Exiting
Waitress writes bill
Waitress goes over to customer
Waitress gives bill to customer
Customer gives tip to waitress
Customer goes to cashier
Customer gives money to cashier
Customer leaves restaurant
From Schank & Abelson (1977). Reprinted
by permission of the publisher. © 1977
by Erlbaum.
TABLE 5.2
Empirical Script Norms at Three Agreement
Levels
Open doora
Enterb
Give reservation name
Wait to be seated
Go to table
Sit downc
Order drinks
Put napkins on lap
Look at menu
Discuss menu
Order meal
Talk
Drink water
Eat salad or soup
Meal arrives
Eat food
Finish meal
Order dessert
Eat dessert
Ask for bill
Bill arrives
Pay bill
Leave tip
Get coats
Leave
aRoman type indicates items listed by at least 25% of the participants.
bItalic type indicates items listed by at least 48% of the participants.
cBoldface type indicates items listed by at least 73% of the participants.
After Bower, Black, & Turner (1979). Adapted by permission of the
publisher. © 1979 by Cognitive Psychology.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 139
various statements came from the story (in another experiment). When recalling
these stories, participants tended to report statements that were parts of the
script but that had not been presented as parts of the stories. Similarly, in the
recognition test, participants thought they had studied script items that had not
actually been in the stories. However, participants showed a greater tendency to
recall actual items from the stories or to recognize actual items than to misrecognize
foils that were not in the stories, despite the distortion in the direction of
the general schema.
In another experiment, these investigators read to participants stories composed
of 12 prototypical actions in an episode; 8 of the actions occurred in
their standard temporal position, but 4 were rearranged. Thus, in the restaurant
story, the bill might be paid at the beginning and the menu read at the end. In
recalling these stories, participants showed a strong tendency to put the events
back into their normal order. In fact, about half of the statements were put
back. This experiment serves as another demonstration of the powerful effect
of general schemas on memory for stories.
These experiments indicate that new events are encoded with respect to general
schemas and that subsequent recall is influenced by the schemas. I have talked
about these effects as if participants were misrecalling the stories. However, it is
not clear that these results should be classified as acts of misrecall. Normally, if a
certain standard event, such as paying a check, is omitted in a story, we are supposed
to assume it occurred. Similarly, if the storyteller says the check was paid at
the beginning of the restaurant episode, we have some reason to doubt the storyteller.
Scripts or schemas exist because they encode the predominant sequence of
events in a particular kind of situation. Thus, they can serve as valuable bases for
predicting missing information and for correcting errors in information.
Scripts are event schemas that people use to reason about prototypical events.
Abstraction versus Exemplar Theories
We have already described semantic networks and schemas as two ways of representing
conceptual knowledge. It is fair to say that although each has merits,
the field of cognitive psychology has concluded that both are inadequate. We
already noted that semantic networks do not capture the graded character of
categorical knowledge such that different instances are better or worse members
of a category. Schemas can do this, but it has never been clear in detail how
to relate them to behavior. The field is currently struggling between two alternative
ways of theorizing about conceptual knowledge. One type of theory
holds that we have actually abstracted general properties from the instances we
have studied; the other type holds that we actually store only specific instances,
with the more general inferences emerging from these instances. We will call
these the abstraction theories and the exemplar theories. The debate between
these two perspectives has been with us for centuries—for instance, in the
debate between the British philosophers John Locke and George Berkeley.
Locke claimed that he had an abstract idea of a triangle that was neither
oblique or right-angled, neither equilateral, isosceles, or scalene, but all of these
140 | Representation of Knowledge
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 140
Conceptual Knowledge | 141
at once, while Berkeley claimed it was simply impossible for himself to have an
idea of a triangle that was not the idea of some specific triangle.
The schema theory we have considered is an abstraction theory, but others
of this type have been more successful. One alternative assumes that people
store a single prototype of what an instance of the category is like and judge
specific instances in terms of their similarity to that prototype (e.g., Reed,
1972). Other models assume that participants store a representation that also
encodes some idea of the allowable variation around the prototype (e.g., Hayes-
Roth & Hayes-Roth, 1977; Anderson, 1991).
Exemplar theories could not be more different. They hold that we store no
central concept, but only specific instances. When it comes time to judge how
typical a specific object is of birds in general, we compare it to specific birds
and make some sort of judgment of average difference. Exemplar theories
include those of Medin and Schaffer (1978) and Nosofsky (1986).
Given that abstraction and exemplar theories differ so greatly in what they
propose the mind does, it is surprising that they generate such similar predictions
over a wide range of experiments. For instance, both types predict better
processing of central members of a category. Abstraction theories predict this
because central instances are more similar to the abstract representation of the
concept. Exemplar theories predict this because central instances will be more
similar, on average, to other instances of a category.
There appear to be subtle differences between the predictions of the two
types of theories, however. Exemplar theories predict that people should be
influenced by studying specific instances similar to a test instance and that such
influences should go beyond any effect of some representation of the central
tendency. Thus, although we may think that dogs in general bark, we may have
experienced a peculiar-looking dog that did not, and we would then tend to
expect that another similar-looking dog would also not bark. Such effects of specific
instances can be found in some experiments (e.g., Medin & Schaffer, 1978;
Nosofsky, 1991). On the other hand, some research has shown that people will
infer tendencies that are not in the specific instances (Elio & Anderson, 1981).
For example, if one has encountered many dogs that chase balls and many dogs
that bark at the postman, one might consider a dog that both chases balls and
barks at the postman to be particularly typical. However, we may never have
observed any specific dog both chasing balls and barking at the postman.
Much of the past research on categorization has been an effort to determine
whether abstraction theories or exemplar theories are correct. The recent trend,
however, has been a recognition that people may sometimes use abstractions
and other times use instances to represent categories (Anderson & Betz, 2001;
Ashby, Alfonso-Reese, Turken, & Waldron, 1998; Erickson & Kruschke, 1998;
Gobet, Richman, Staszewski, & Simon, 1997; Palmeri & Johansen, 1999; Smith &
Minda, 1998). Perhaps the clearest evidence for this expanded view comes from
neuroimaging studies showing that different participants use different brain
regions to categorize instances. For example, Smith, Patalano, and Jonides
(1998) had participants learn to classify a set of 10 animals like those shown in
Figure 5.14. One group was encouraged to use rules such as “An animal is from
Venus if at least three of the following are true: antennae ears, curly tail, hoofed
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 141
feet, beak, and long neck. Otherwise it is from Saturn.” Participants in a second
group were encouraged simply to memorize the categories for the 10 animals.
Smith et al. found very different patterns of brain activation as participants
classified the stimuli. Regions in the prefrontal cortex tended to be activated in the
abstract-rule participants, whereas regions in the occipital visual areas and the
cerebellum were activated in the participants who memorized instances. Smith
and Grossman (in press) review evidence that this second exemplar system also involves
brain regions supporting memory such as the hippocampus (see Figure 1.7).
There may be multiple different ways of representing concepts as abstractions.
Although the Smith et al. study identified an abstract system that involves
explicit reasoning by means of rules, there is also evidence for abstract systems
that involve unconscious pattern recognition—for instance, our ability to distinguish
dogs from cats, without being able to articulate any of the features that
separate the two species. Ashby and Maddox (2005) argue that this system depends
on the basal ganglia (see Figure 1.8). As they review, damage to the basal
ganglia (as happens with Parkinson’s and Huntington’s disease) results in
deficits in learning such categories. The basal ganglia region has been found to
be activated in a number of studies of implicit category learning.
Categories can be represented either by abstracting their central tendencies
or by storing many specific instances of categories.
Natural Categories in the Brain
The studies we have been discussing look at the learning of new laboratorydefined
categories. There has always been some suspicion about how similar
such laboratory-defined categories are to the kinds of natural categories that
we have acquired through experience, such as birds or chairs. Laboratory categories
display the same sort of fuzzy boundaries that natural categories do
and share a number of other attributes. However, natural categories arise over
a much longer time period than a typical laboratory task. Over their long
learning history, people come to develop biases about such natural categories
142 | Representation of Knowledge
FIGURE 5.14 Examples of the drawings of artificial animals used in the PET studies of
Smith, Palatino, and Jonides showing that people sometimes use rule-based abstractions
and sometimes use memory-based instances to represent categories. (After Smith, Palatino, & Jonides,
1998. Adapted by permission of the publisher. © 1998 by Cognition.)
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 142
Conceptual Knowledge | 143
as living things and artifacts.Much of the research documenting these biases has
been done with primary-school children who are still learning their categories
but have advanced beyond younger children. For instance, if primary-school
children are told that a human has a spleen, they will conclude that dogs have
spleens too (Carey, 1985). Similarly, if they are told that a red apple has pectin
inside, they will assume that green apples also have pectin (Gelman, 1988).
Apparently, children assume that if something is a part of a member of a
biological category, it is an inherent part of all members of the category. On the
other hand, if told that an artifact such as a cup is made of ceramic, they do not
believe that all cups are made of ceramic. The pattern is just the opposite with
respect to actions. For instance, if told that a cup is used for “imbibing” (a term
they do not know), they believe that all cups are used for imbibing. In contrast,
if told that they can “repast” with a particular red apple, they do not necessarily
believe that they can repast with a green apple. Thus, artifacts seem distinguished
by the fact that there are actions appropriate to the whole category of
artifacts. In summary, children come to believe that all things in a biological
category have the same parts (like pectin in apples) but all things in an artifact
category have the same function (like imbibing for cups).
Cognitive neuroscience data suggest that there are different representations
of biological and artifact categories in the brain. Much of this evidence comes
from patients with semantic dementia, who suffer deficits in their categorical
knowledge because of brain damage. Patients with damage to different regions
show different deficits. Patients who have damage to the temporal lobes suffer
deficits in their knowledge about biological categories such as animals, fruits,
and vegetables (Warrington & Shallice, 1984; Saffran & Schwartz, 1994). These
patients are unable to recognize such objects as ducks, and when one was asked
what a duck is, the patient was only able to say “an animal.”However, knowledge
about artifacts such as tools and furniture is relatively unaffected in these
patients. On the other hand, patients with frontoparietal lesions are impaired
in their processing of artifacts but unaffected in their processing of biological
categories.
Table 5.3 compares example descriptions of biological categories and artifact
categories by two patients with temporal lobe damage. These types of
patients are more common than patients with deficits in their knowledge
of artifacts.
It has been suggested (e.g.,Warrington & Shallice, 1984; Farah & McClelland,
1991) that these dissociations occur because biological categories are more associated
with perceptual categories such as shape, whereas artifacts are more
associated with the actions that we perform with them. Farah and McClelland
offer a computer simulation model of this dissociation that learns associations
among words, pictures, visual semantic features, and functional semantic features.
By selectively damaging the visual features in their computer simulation,
they were able to produce a deficit in knowledge of living things; and by selectively
damaging the functional features, they were able to produce a deficit in
knowledge of artifacts. Thus, loss of categorical information in such patients
seems related to loss of the feature information that defines these categories.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 143
Recent brain-imaging data also seem consistent with this conclusion (see
A. Martin, 2001, for review). In particular, it has been shown that when people
process pictures of artifacts or words denoting artifacts, the same regions of
the brain that have been shown to produce category-specific deficits when damaged
tend to be activated. Processing of both animals and tools activates regions
of the temporal cortex, but the tool regions tend to be located above (superior
to) the animal regions. There is also activation of occipital regions (visual
cortex) when processing animals. In general, the evidence seems to point to a
greater visual involvement in the representation of animals and a greater motor
involvement in the representation of artifacts. There is some debate in the literature
over whether the real distinction is between natural categories and artifacts
or between visual-based and motor-based categories (Caramazza, 2000).
There are differences in the way people think about biological categories and
artifact categories and differences in the brain regions that support these
two types of categories.
•Conclusions
We remember only a tiny fraction of what we experience. Estimates of the
storage capacity (e.g., Treves & Rolls, 1994; Moll & Miikkulainen, 1997) of the
brain differ substantially, but they are all many orders of magnitude less than
144 | Representation of Knowledge
TABLE 5.3
Performance of Two Patients with Impaired Knowledge of Living Things
on Definitions Task
Patient Living Things Artifacts
1 Parrot: Don’t know Tent: Temporary outhouse, living home
Daffodil: Plant Briefcase: Small case used by students
Snail: An insect animal to carry papers
Eel: Not well Compass: Tool for telling direction you
Ostrich: Unusual are going
Torch: Handheld light
Dustbin: Bin for putting rubbish in
2 Duck: An animal Wheelbarrow: Object used by people
Wasp: Bird that flies to take material about
Crocus: Rubbish material Towel: Material used to dry people
Holly: What you drink Pram: Used to carry people, with wheels
Spider: A person looking and a thing to sit on
for things, he was a spider Submarine: Ship that goes underneath
for his nation or country the sea
After Farah & McClelland (1991). Adapted by permission of the publisher. © 1991 by Journal of Experimental
Psychology: General.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 144
Key Terms | 145
1. Jill Price, a person with superior autobiographical
memory described at the beginning of the chapter,
can remember what happened on almost any day of
her life (see her interview with Diane Sawyers: http://
abcnews.go.com/Health/story?id=4813052&page=1).
For instance, if you ask her, she can tell you the date
of the last show of any former TV series she watched.
On the other hand, she reported great difficulty in
remembering the dates in history class.Why do you
think this is?
2. Take some sentences at random from this book and try
to develop propositional representations for them.
3. Barsalou (2008) claims little empirical evidence has
been accumulated to support amodal symbol systems.
What research reviewed in this chapter might be
considered evidence for amodal symbol systems?
4. Consider the debate between amodal theories and
multimodal theories and the debate between exemplar
and abstraction theories. In what ways are these
debates similar and in what ways are they different?
Questions for Thought
Key Terms
abstraction theory
amodal hypothesis
amodal symbol system
arguments
default values
dual-code theory
embodied cognition
exemplar theory
isa link
link
mirror neuron
mnemonic technique
multimodal hypothesis
node
perceptual symbol
system
proposition
propositional network
propositional
representation
relations
schema
script
slot
what would be required to store a faithful video recording of our whole life.
This chapter has reviewed the studies of what we retain and what we forget—
for instance, what subject was being taught, but not what the teacher was
wearing (Figure 5.3), or that we were in an office, but not what was in the
office (Figure 5.11). The chapter also reviewed three perspectives on the basis
for this selective memory.
1. The multimodal hypothesis (Figure 5.9a) that we select aspects of our
experience to remember and often convert from one medium to another.
For instance, we may describe a room (visual) as an “office” (verbal).
This hypothesis holds that we maintain the perceptual-motor aspects
of our experience but only the significant aspects.
2. The amodal hypothesis (Figure 5.9b) that we convert our experience
into some abstract representation that just encodes what is important.
For instance, the chapter discussed how propositional networks
(e.g., Figure 5.7) captured the connections among the concepts in
our understanding of a sentence.
3. The schema hypothesis that we remember our experiences in terms of
the categories that they seem to exemplify. These categories can either
be formed from a bundle of specific experiences or an abstraction like
a category.
These hypotheses need not be mutually exclusive, and cognitive scientists are
actively engaged in trying to understand how to coordinate these explanations.
Anderson7e_Chapter_05.qxd 8/20/09 9:44 AM Page 145
146
6
Human Memory: Encoding
and Storage
Past chapters have discussed how we perceive and encode what is in our present.
Now we turn to discussing memory, which is the means by which we can
perceive our past. People who lose the ability to create new memories become
effectively blind to their past. I would recommend the movie Memento as providing
a striking characterization of what it would like to have no memory. The protagonist
of the film, Leonard, has anterograde amnesia, a condition that prevents him from
forming new memories. He can remember his past up to the point of a terrible
crime that left him with amnesia, and he can keep track of what is in the immediate
present, but as soon as his attention is drawn to something else, he forgets what
has just happened. So, for instance, he is constantly meeting people he has met
before, who have often manipulated him, but he does not remember them, nor can
he protect himself from being manipulated further. Although Leonard incorrectly
labels his condition as having no short-term memory, this movie is an accurate
portrayal of anterograde amnesia—the inability to form new long-term memories. It
focuses on the amazing ways Leonard tries to connect the past with the immediate
present.
This chapter and the next can be thought of as being about what worked and
did not work for Leonard. This chapter will answer the following questions: • How do we maintain a short-term or working memory of what just happened?
This is what still worked for Leonard. • How does the information we are currently maintaining in working memory
prime knowledge in our long-term memory? • How do we create permanent memories of our experiences? This is what did
not work any more for Leonard. • What factors influence our success in creating new memories?
Anderson7e_Chapter_