JOURNAL OF MATHEMATICAL PSYCHOLOGY 13,243-268 (1976)

Book Review

Subproblems of Semantic Memory:

A Review of “ Human Associative Memory” by
J. R. Anderson and G. H. Bower*?+

WAYNE A. WICKELGREN

Lkpartment of Psychology, University of Oregon, Eugene, Oregon 94703

In the first place, I liked the title; human conceptual memory is associative, and it is
nice to see that assertion right up front in the title of this important book. I have always
considered the issue of whether any memory system was associative or nonassociative
to be fundamental. The first memory paper I ever wrote (Wickelgren, 1965) presented
a definition of one type of associative memory and one type of nonassociative memory
and asserted that the evidence favored the assumption that verbal short-term memory
is associative. Since then, about half a dozen subsequent papers of mine were also
concerned with this issue and related issues concerned with the associative character of
verbal long-term memory (Wickelgren, 1972) and the nonassociative character of
visual sensory memory (Wickelgren & Whitman, 1970).
Partly because the general notion of associative memory has been around since
Aristotle, the most common reaction to my work on this topic has been largely “ho
hum.” Also, the decade of the 60’s was not a good time for definitions, arguments, and
evidence favorable to the assumption that human conceptual memory is associative.
Computer models and transformational generative grammar were “in,” and, for
completely illogical reasons, associative memory was “out.” The notion that memory
is associative suffered from “guilt by association.” Nearly universal acceptance of the
hypothesis of associative memory throughout the early history of psychology led to the
ridiculous notion during the 60’s that the assumption of associative memory was
directly linked to a noncognitive, operationist, crassly empirical, S-R approach to
psychology. As a consequence, a lot of psycholoinguists and cognitive memory
researchers threw out the baby with the bath water and went off developing theories
of the human mind that either ignored memory altogether or worked with some

* This work was supported by Grant 3-0097 from the National Institute of Education and by
contract F 44620-73-C-0056 from the Advanced Research Projects Agency, D. 0. D. to Ray
Hyman.
+ New York: John Wiley and Sons, New York, 1973.
243
Copyright Q 1976 by Academic Press, Inc.
AU rights of reproduction in any form reserved.
244 WAYNE A. WICKELGREN

primitive and incorrect nonassociative conception of it. Research of the last few years
on semantic memory, particularly as exemplified by the Anderson and Bower book,
has gone a long way toward rectifying this grievous error.
In the second place, I liked the overview that Anderson and Bower provided to the
different aspects of understanding semantic memory. As I see it, a complete theory of
semantic memory must solve seven theoretical subproblems: the structure of semantic
memory, recognition, acquisition, storage, retrieval, inference, and generation. The
structure of semantic memory refers to the nature of the memory code, including its
associative vs. nonassociative nature, the particular types of associations and elements
(nodes), how our knowledge of the world is represented in semantic memory, etc.
Recognition refers to the process by which the familiar portion of new stimulus input is
identified and mapped onto already existing representation in semantic memory,
including its dynamics. Acquisition refers to the process by which the new (previously
uncoded) portion of input is encoded into sematic memory, including its dynamics.
Storage referes to the nature and dynamics of the changes that take place in semantic
memory over the retention interval between acquisition and usage (retrieval).
Retrieval, inference, and generation could all be subsumed under the rubric of
“usage,” but the theoretical problems involved in each seem sufficiently different at
the present time to warrant separate treatment. Accordingly, I use the term “retrieval”
in the same way that Anderson and Bower use the term “fact retrieval” to refer to
elementary question answering (which differs from Anderson and Bower’s use of the
term “question-answering”). Elementary question answering has two principal
components: answering yes-no questions and answering wh-questions (who, what,
when, where, and possibly why, and how). This subdivision corresponds approximately
to the distinction in traditional memory research between recognition and recall
(approximately, but not exactly). “Retrieval” will be used to refer to elementary
question answering for material directly stored in semantic memory. By contrast,
inference will refer to elementary question answering that requires some combination
of separately stored memory traces in semantic memory (e.g., if Liebnitz was a person
and people have four-chambered hearts, then Liebnitz had a four-chambered heart).
Generation refers to spoken or written verbal production (principally speech produc-
tion). This is distinguishable from elementary retrieval and inference by virtue of the
production of long, syntactically complex utterances.
The processes are not totally independent. For example, recognition is usually
involved in acquisition and necessarily in usage. The last six subproblems refer to
processes for which one wants ultimately to specify both the basic nature of the process
and its (temporal) dynamics. The structure of semantic memory has no dynamic aspect
and is perhaps the most basic subproblem to solve, since it may be deeply involved in
the solution to most, if not all, of the other theoretical subproblems. However, studying
the structure of semantic memory typically requires at least some minimal assumptions
concerning most of the six process. Despite the likelihood of at least a modest degree
 SUBPROBLEMS OF SEMANTIC MEMORY 245

of interaction between each of these seven subproblems, some such analysis is essential
to the solution of any complex scientific problem. This seems like a good one.
One of the strengths of the Anderson and Bower book is the presentation of a sub-
problem analysis which seems to me to be equivalent to the one just specified, or
nearly so. Furthermore, Anderson and Bower present preliminary examples of a
theoretical solution to each of the first five subproblems and at least discuss the problem
of an adequate accounting for inference in some detail, although their theory of
semantic memory has very limited inferential capacity. Anderson and Bower mention
the problem of speech generation, but present no example theoretical solution.
Anderson and Bower generally also discuss a number of empirical findings (their own
and others) that confirm or reject various aspects of their theory. In the scope of both
their theoretical approach to semantic memory and the empirical testing thereof, the
Anderson and Bower theory (HAM for human associative memory) has no published
rival. It should be noted, that this does not imply that Anderson and Bower’s theoretical
solution to any subproblem area is superior to that of Norman and Rumelhart (1975),
Schank (1973), Winograd (1972), etc., to name just a few outstanding recent alternative
formulations. The ultimate worth of different theories, particularly in the most basic
area of the structure of semantic memory will take some time to determine. I certainly
am not able to draw any definite global conclusion on this matter at the present.
Before proceeding to a critique of Anderson and Bower’s theory, I wish to provide
a brief overview of the book’s contents and evaluate its suitability for courses. The
book is divided into two parts. The first five chapters are an historical introduction to
semantic memory including various versions of associationism, Gestalt and recon-
structive notions of memory (such as they are), computer simulation or artificial
intelligence models of semantic memory, and linguistic contributions to the under-
standing of syntax and semantics. This introductory review occupies the first 135 pages
and substantially enhances the value of the book as a text for a semantic memory
seminar or an advanced course.
The remaining three quarters of the book presents Anderson and Bower’s theory
with chapters having some rough correspondence to the subproblem analysis previously
described, though this could have been improved to give a more exact correspondence.
In addition, Anderson and Bower extend their theory to the analysis of the syntactically
unstructured materials used in traditional verbal learning experiments. Finally, at
various stages in the book, Anderson and Bower presented some ideas concerning the
propositional representation of image memory. This is very stimulating, whether
or not you agree with their point of view.
Doug Hintzman, Ray Hyman, and I used the Anderson and Bower book as the sole
text for a one quarter graduate seminar on semantic memory, and in my opinion it was
an unqualified success. However, many’students, particularly those with little or no
background in linguistics, psycholinguistics, or artificial intelligence complained that
the book was too difficult to understand. Personally, I think that is a matter of expecta-
246 WAYNE A. WICKELGREN

tion. Students who are familiar with courses and seminars that are focused on
experiments will find a course that is primarily oriented to the understanding of a
relatively complex and precisely specified theory to be conceptually difficult. If students
expect such material to be difficult, to require substantial time for understanding and
to require asking a lot of questions, then a course or seminar based on this book
can be completely successful. Despite complaints from a number of students, the
semantic memory seminar we held based on this book has had an enormous impact
at the University of Oregon, certainly greater than any course or seminar I have ever
been a part of before.
In my opinion, qualitative and quantitative mathematical psychology have reached
the point of development where we should insist that the vast majority of all students
in “experimental” psychology develop some conceptual (mathematical) sophistication
to accompany their experimental and statistical sophistication. Since semantic memory
is currently “hot,” systematic exposure to this book provides an excellent vehicle for
increasing the conceptual and theoretical training of graduate students in cognitive
psychology. I do not think that the proper way to introduce the area of semantic
memory is by prior study of linguistics or computer science, since these areas provide
less motivation to the student in the form of the psychological relevance of the
conceptual material being learned. Finally, prior study of empirical psycholinguistics
is of only minor value in understanding the theoretical conceptions important in
semantic memory and simply lengthens the time to acquire the necessary sophistication.

STRUCTURE OF SEMANTIC MEMORY

Associative Memory

As in any associative memory, Anderson and Bower’s semantic memory is composed

of nodes and links (associations). Anderson and Bower develop the notion that there
is more than one type of link connecting the nodes of semantic memory, and they have
a specific proposal regarding the different types of labeled links. However, their
proposal may be somewhat open-ended in that they are concerned with only a simple
subset of all English sentences. The propositions that can be represented in HAM
have five principal constituents (subject, verb, object, location, and time). Each of these
constituents may be modified by an adjective or a dependent proposition. Yes-no and
wh-questions may be asked concerning any of the five constituents. Anderson and
Bower restricted the form of the sentences that could be input to semantic memory.
It was not clear to me whether Anderson and Bower thought that the set of labeled
links which they employed in dealing with this subset of English sentences would be
sufficient to handle the underlying semantic (propositional) structure of all sentences
that can be encoded into semantic memory. As currently formulated, HAM consists of
 SUBPROBLEMS OF SEMANTIC MEMORY 247

approximately 30 different types of associations organized in pairs of mutual inverses

(forward and backward associations).
In my simplistic intuitive neurophysiologizing, it seems somewhat implausible to
imagine that there are say 30 chemically distinct types of synapses. However, even if
there were only one type of synaptic connection in the area of the brain concerned with
semantic memory, for the purposes of an abstract psychological model, it might well be
desirable to assume 30 different types of links, especially since we do not know whether
psychological links correspond to sets of physiological synapses (though this has always
been the primary hypothesis).

Propositional Nodes

Besides having a multiplicity of labeled links, Anderson and Bower have taken
another important step away from traditional associative memory by assuming that
learning consists not of the formation or facilitation of connections between already
existing nodes (ideas, concepts), but rather “vertical” associations between existing
nodes for lower-level concepts, phrases or propositions and higher-level nodes
representing phrases and propositions. For example, to encode the phrase “touch
debutante,” Anderson and Bower do not assume the formation or facilitation of a
single link between the concept “touch” and the concept “debutante.” Rather, they
assume that a new node is formed standing for the predicate “touch debutante” to
which both “touch” and “debutante” are linked in both forward and backward
directions. Similarly, the proposition “hippie touch debutante” is encoded in semantic
memory, not by direct associations between each of the pairs of concepts, but rather
by a hierarchy of “vertical” associations: (a) “touch” and “debutante” to the predicate
nodes “touch debutante” and (b) the predicate node and the subject “hippie” to the
propositional node “hippie touch debutante.”
To me this assumption of the existence of phrasal and propositional nodes was the
single most important idea in the book. Anderson and Bower’s theory does not quite
assume that phrasal and propositional nodes are equivalent to more elementary concept
nodes (and I consider this to be a defect of the theory). However, Anderson and
Bower’s theory deviates from the traditional assumption that associations are
“horizontal” (directly connecting the component ideas of a proposition) more than the
theories of either Norman and Rumelhart (1975), or Schank (1974).
The assumption that associative memory requires the capability of adding new nodes
by means of these vertical associations is an idea that seems very attractive and
necessary to me. For example, I have argued elsewhere (Wickelgren, 1969) in favor
of the assumption that associative memory can add new nodes or elements to stand
for new concepts or ideas as they are learned. Horizontal associations between the
attributes of a concept seem completely insufficient to explain our capacity for
representation of concepts. Also, I once used a somewhat similar notion of vertical
248 WAYNE A. WICKELGREN

associations of items to serial position concepts to explain rehearsal grouping

phenomena in short-term memory (Wickelgren, 1964, 1967). Miller’s (1956) famous
chunks and Estes’ (1972) control elements also assume that component items can be
linked together by vertical associations to a new superordinate element. Finally, as
Anderson and Bower suggest, it is possible to interpret some aspects of the Gestalt and
reconstructionist theories of memory as involving the assumption that nodes can be
added to associative memory. So there are predecessors of this basic idea, but to my
knowledge no one before ever proposed the elegant idea that new nodes might be
introduced in memory to represent phrases and propositions. Although Anderson and
Bower’s theory does not quite propose this, it proposes something so close to it that,
should this idea prove to be useful, it must rank as a major contribution by Anderson
and Bower to cognitive psychology.
Actually, Anderson and Bower economize on the number of nodes in memory by
representing conjunctive compounds of phrases and propositions with single nodes as
often as possible in preference to unique node representation of each phrase or
proposition. For example, if HAM has already stored that “John hit the ball” and
later is presented with “John touched the debutante,” it will link the new predicate to
the old propositional node in such a way as to form a propositional node representing
the compound sentence “John hit the ball and touched the debutante.” Similarly, at
the phrasal level, if HAM has already stored in memory that “a hippie touched a
debutante” and later receives the sentence that “the same hippie kissed the same
debutante,” HAM will link “touched” and “kissed” to the same predicate node to
form a compound predicate representing “kissed and touched the debutante.”
This appears to be a mistake. Individual propositions should have separate nodal
representation. One reason for this is to permit easy encoding of a subsequent contra-
diction to one element of the compound proposition. As I understand it, HAM’s
current method of representing propositional negation, in combination with single
node representation of compound propositions, would not permit one to later con-
tradict a single propositional component of the compound. HAM encodes the negation
of an existing proposition by embedding the proposition into the higher order node
“It is false that . ..” HAM’s treatment of propositional negation seems correct. The
assumption of compound phrasal and propositional nodes seems incorrect.
What appears to me to be a more reasonable way to represent a compound sentence
such as “the hippie touched the debutante and the policeman touched the debutante”
is to assume separate propositional and subject nodes, but links to a common predicate
node (“touched the debutante”). An associative semantic memory must recognize
the commonality of two predicates, but not at the price of inability to later inde-
pendently negate either proposition, to say nothing of the ability to independently
retrieve one proposition without the other. Graphical illustration of the difference
between these two modes of representing compound propositions is shown in
Figure 1.
 SUBPROBLEMS OF SEMANTIC MEMORY 249

policemar\hippie-/ouched debutante
FIG. 1A. Anderson and Bower’s representation of a conjunctive compound proposition.
S = Subject, P = Predicate, R = Relation (verb), 0 = Object.
FIG. 1B. Independent propositional representation of a compound proposition.

Semantic Memory is a Network

In agreement with all psychological and computer science models of semantic
memory with which I am familiar, HAM represents memory by means of a network,
rather than a set of independent hierarchies such as were commonly found representing
propositional material in the early days of the Chomskian revolution in linguistics.
I suppose no one ever believed that semantic memory was a set of independent
hierarchies, but it is worth pointing out that in “deep structure” (here interpreted to be
equivalent to semantic memory), as in surface structure, one proposition can be
embedded as a constituent of another proposition. The same node for a concept or
phrase should be considered as linked into every propositional structure in semantic
memory in which it occurs.

Between-Propositional Analysis
One of the many useful distinctions made in Anderson and Bower’s book is the
distinction concerning the between-propositional level of analysis and the within-
propositional level of analysis. The distinction has some analogy to that between
propositional and predicate calculus, though Anderson and Bower’s within-proposi-
tional constituent structure is far more detailed than the constituent structure of
predicate calculus.
Basically, Anderson and Bower’s theory of the between-propositional level of
analysis simply states that compound and complex sentences have as their constituents
the propositions represented by the various component clauses of the sentence.
I guess everyone in linguistics and psycholinguistics agrees with this, and, in
Chapter 9, Anderson and Bower present some evidence supporting the psychological
reality of this between-propositional level of analysis in the form of positive transfer
for learning new sentences that contain previously learned complete subject + verb +
object propositions.

Adjectives. Somewhat more debatable is the assumption made by Anderson and

Bower (in agreement with Chomsky) that the semantic memory (deep structure)
representation of adjectives is equivalent to that of a modifying dependent clause. That
250 WAYNE A. WICKELGREN

is to say, a sentence such as “John bit a red apple” would be represented by two
constituent propositions: “John bit an apple” and “the apple is red.” Although
Anderson and Bower appear to assume this deep structure analysis of adjectives, they
do not do experiments using adjectives and appear to have largely ignored the issue.
I do not remember the exact experiments that have been performed on this question,
but it is my impression that there is precious little psychological evidence that supports
the separate propositional representation of adjectives. Anderson and Bower might
handle adjective and noun phrases in a manner similar to their treatment of definite
descriptions, but the success of such treatment cannot presently be evaluated.

Within-Propositional Representation

In my opinion, the heart of any theory of semanticmemory is its representationof

the components of a proposition. There are quite a variety of different proposals
regarding within-propositional representation, including the theory of Anderson and
Bower, but little empirical support for one over another and predictive inadequacies
for all of them.
Although Anderson and Bower admit (p. 509) that their within-propositional
representation hasvery little support, it is still interesting to consider it. In agreement
with Chomsky, Anderson and Bower assumea predicate grammar that conjoins the
verb (relation) and object together to form a predicate node and then conjoins the
predicate node to the subject (or agent) to form the fact node.

Context, events, scenarios. Th e somewhat novel and interesting component of

Anderson and Bower’s within-propositional representationis the existenceof a context
node which is conjoined with the fact node to form the high level propositional node.
The immediate constituents of the context node are location and time. Thus, instead
of location and time being modifiers of the verb or constituents of the predicate, they
are high level constituents of the proposition on a par with the (nuclear) fact node. In
HAM all of the facts that occur in a common context are linked to the samecontext
node, which implies that the context node might be consideredto represent an event
(although temporal sequenceswould not be represented). In HAM, such event
representation is done by linking the multiple fact nodes to the same high level
propositional node creating a compound propositional node. However, for the reasons
stated previously, it seemssuperior to create new propositional nodes for each of the
facts that occur in a commoncontext, but link eachto the commoncontext node. Thus,
either the highest level node in HAM or the context node in the presently proposed
theory representsthe event (set of facts occurring in a common context). This seems
intuitively nice, though no way has been specified here to represent the sequenceof
facts or events (a scenario) and representation of this might introduce considerable
complexity.
 SUBPROBLEMS OF SEMANTIC MEMORY 251

Experiments. The experiment in Chapter 9 that demonstrated positive transfer

in the learning of new propositions containing a previously learned fact constituent
(subject + verb + object proposition) also demonstratedno positive transfer in the
caseof previously learned subject + verb, verb + object (predicate) or subject +
object constituents. Anderson and Bower’s theory and any theory that assertsa reality
to the predicate node would seemingly require positive transfer for the verb + object
constituent in comparisonto a no-repetition control condition or the subject + verb or
subject + object repetition conditions. Casegrammarsof the type used in linguistics
by Fillmore (1968) and in semantic memory by Rumelhart, Lindsay, and Norman
(1972) might predict the verb to be closely related to both the subject and the object,
obtaining positive transfer for both of these cases,but not for the subject + object
repetition. However, no positive transfer was obtained for any of these pairs of
constituents, which I suspectindicates that something is basically wrong with both of
thesebasic conceptionsof the constituent structure of propositions.
One possibleway out is to assumethat the fact node decomposesinto subject, verb,
and object constituents without any intermediate node structure. However, this
suggeststhat the subject and object are as closely related to each other asthe subject
and the verb or the verb and the object. This is contradicted by everyone’s structural
intuition. Since positive transfer was obtained for repetition of an entire fact, positive
transfer experiments of this type appearto be relevant for evaluating the nodal structure
of semantic memory. Hence, we are faced with a puzzling theoretical problem, the
solution to which may go a long way toward adequatewithin-propositional representa-
tion.

Being verbs. A minor elegant feature in the Anderson and Bower theory is the fact
that verbs of being (is, are, was, were, etc.,) are not representedin the sameway as
other verbs, but rather are consideredto be representedby an undifferentiated predicate
node as shown in Figure 2. The difference in the representation of a sentencewith
a verb of being comparedto one with a transitive verb is shown in Fig. 1. It should be
noted that the representationsshown in Figures I and 2 are simplified by deletion of
terminal quantifiers from the representation assumedby HAM.

in the park.”
6d P&PI.
Presed

FIG. 2. Representations
Canary

C = context,
icate, R = Relation (Verb),
in HAM
tr ar

of “A canary is a bird”
F = Fact, L = Location,
and 0 = Object.
re en
. I
9
and “A canary is biting the man
T = Time, S = Subject, P = Pred-
252 WAYNE A. WICKELGREN

Subjectsor agents. Anderson and Bower waffle considerably concerning whether

they prefer to use deep subjects or agents. The distinction primarily arises concerning
the question of the representation of passive sentences. The subject of a passive
sentence is the object of the corresponding active sentence and vice versa. One could
assume that passive sentences are represented with the subjects and objects reversed
compared to the representation of the corresponding active sentences. However, it
has appeared parsimonious to many people including Anderson and Bower to assume
instead an underlying unitary representation for both passive and active forms of the
sentence, replacing the notion of subject with the notion of agent. A similar solution
might be taken for the representation of “John is taller than Bill” vs. “Bill is shorter
than John,” where we might consider the marked form (“Bill is shorter than John”)
to be represented using the unmarked relation (“John is taller than Bill”).
Unfortunately, experiments performed by Anderson and Bower do not
unambiguously support the notion of a common underlying form of encoding for
passive and active sentences, though they do suggest that such sentences are closely
related (as one feels sure must be the case). The need to explain why passive sentences
exist in the language in combination with the desirability of simply representing the
equivalence of active and passive forms is another important unsolved problem.

Structural vs. temporal contiguity. The sad state of affairs regarding within-
propositional representation is nowhere more convincingly demonstrated than in
pages 319-329 where physical (temporal) contiguity sometimes provides better
prediction than HAM of what concepts are most closely associated to other concepts
within propositions. Only a fool could think that the physically adjacent words in a
sentence constitute the principal structure of semantic memory. Hence, I interpret the
failure of HAM’s propositional contiguity by comparison to temporal contiguity to
reflect the miserably inadequate state of our understanding concerning within-proposi-
tional representation. This is especially so considering that, once again, the analysis
of between-propositional structure was quite superior to physical contiguity in
predicting what is most closely associated to what in sentence recall.

QuantiJiers. Quantifiers are represented in HAM in a very interesting way. Instead

of considering a quantifier (all, some, etc.,) to be a high-level constituent of a proposi-
tion which they appear to be in symbolic logic, Anderson and Bower attach quantifiers
to terminal nodes of propositions. Quantifiers have always been somewhat mysterious
to me and I do not pretend to understand very well the implications of Anderson and
Bower’s decision in favor of terminal quantification. However, the basic idea is that
there is something in common between the concepts for each individual dog that we
know in the world, each subset of dogs, and the concept all dogs. A given proposition
refers to some particular dog or to some particular subset of dogs or to all dogs, but
whichever of these particular concepts is referred to in a proposition, that concept
is conceived to be associated via the appropriate quantifier link to the common concept
 SUBPROBLEMS OF SEMANTIC MEMORY 253

dog. Hence, if your dog, Rover, bit a particular postman there would be a node
representing Rover which was associated via the set membership link to the concept
dog, and there would also be the same type of link between the node representing the
particular postman that he bit and the general concept of postman. Finally, the same
type of link would exist between the node representing the particular act of biting that
your dog performed and the general concept of biting.
The representation of Rover bit all postmen is as shown in Figure 3, where the
particular node for Rover is linked via the set membership association (E) to the
concept of dog and all postmen is linked via the “generic” association for universal
quantification (V) to postman, and the particular act of biting that Rover (or dogs ?) do
(does) to postmen (or all people or no restriction ?) is linked by the generic association
to a more general concept of dog-biting (?), which in turn is linked by the subset
association (C) to the general concept of biting. Of course this sentence is a bit absurd
since it asserts that Rover is currently biting or has bitten or will bite all postmen that
exist. However, I wanted to represent all three types of quantifiers in HAM using a
single diagram. I believe I did this.

do bit es
9 pOStb?lCUl

FIG. 3. Representation of the three types of quantifiers in HAM: member (c), subset (C),
and generic (V meaning “all” or “for all”).

It seemslike a pretty complex structure to attach to the bottom of every terminal

node in a propositional structure. However, this somewhatclumsy looking structure
should not be discarded without considering its desirable properties. For one thing,
there are a large number of different concepts of individual dogs which are to be
distinguished from the concept of all dogs and from the conceptsfor each subset of
dogs (Terriers, Dobermans, etc.,). Yet they all have something in common. HAM
represents this, and, although it complicates the structure somewhatto do so, we
cannot dispensewith this unlesswe can propose a satisfactory alternative representa-
tion. Perhaps it is unnecessary to distinguish between the particular kind of biting that
a dog doesand the particular kind of biting that a lion doesin the manner suggestedby
Anderson and Bower. However, theseare distinct concepts,and it is not clear one way
or the other to me, intuitively, how they should be distinguishedin semanticmemory.
254 WAYNE A. WICKELGREN

I am not fond of the procedure of labeling the links. One alternative is that “Rev-er is
a dog” could be represented by a simple subject-predicate construction. A similar
solution could be used for all subsets of dogs. However, there are important subtleties
here that Anderson and Bower discuss (somewhat opaquely). For example, the Terrier
subset of dogs is not itself a dog. Each member of the subset is a dog, but the subset is
not a dog. The relation between individual elements and subsets of these elements
must be represented. One of the major contributions of the Anderson and Bower
book is to draw attention to this important problem and to suggest one possible
solution.

Concepts and words. Anderson and Bower also draw explicit attention to the
distinction between the concept (idea) and the word representing the concept, that is
between dog and “dog.” The word representing the concept is linked to the general
concept node via the word-idea association. One might try to dispense with unitary
word representatives and consider the representative of a word to be the set of its
phonetic (segmental) constituents. However, I suspect that a unitary (chunk or
concept) representation of words is also necessary.

Types and tokens. Finally, linking terminal constituents of propositional trees via
quantifier associations to more general concepts, in my opinion, exhibits a superb use
of the type-token distinction. Instead of pretending that we create multiple tokens
willy-nilly every time the same concept appears in a new sentence, Anderson and Bower
assume that tokens are introduced only to represent (sometimes subtle) differences in
meaning for the same type. Hence, when the term “the dog” is used in one sentence, it
may refer to Rover, and, in another sentence, it may refer to a different dog. Obviously,
we should represent these by two different concepts. Anderson and Bower introduce
nodes to represent these two quite properly different concepts, linking each of them via
the set membership relation (quantifier) to the more general concept of dog. If both
of these dogs have names, then each individual dog concept would be linked to its
own word representative. However, many concepts have no name of their own and
require disambiguation by an appropriate context for unique identification. The fact
that these concepts do not have single words associated to them does not mean that
the concepts should not have distinct nodes representing them. Anderson and Bower
quite properly decide that they should. Once again, one must be impressed by the deep
sophistication that Anderson and Bower demonstrate in understanding the problems of
adequate representation for semantic memory.
I do not personally believe that every individual act of biting should be represented
by a new node (token), and it is not clear what Anderson and Bower believe regarding
such matters. It seems to me that, at some point, representation in semantic memory
must ignore subtle differences in favor of completely equivalent encoding of similar
concepts in different propositions, allowing the propositional context to further
distinguish the concept in this context.
 SUBPROBLEMS OF SEMANTIC MEMORY 255

Verbal Learning
Anderson and Bower alsoassertthat traditional verbal learning material is encoded
by somewhatdegeneratepropositions. This seemslike a reasonableextension of their
theory to such material and recent work on the efficacy of various mnemonic devices
that in most instances amount to embedding pairs of verbal items into verbal or
imaginal “propositions” is certainly consistentwith this generalpoint of view. Similarly,
the importance of serial position concepts and grouping concepts for the organization
of linear orders, such as in serial list learning, also argues for some more abstract
propositional structure underlying this type of learning as well. In the case of
syntactically unstructured material, such as typically used in verbal learning
experiments, it is not clear that much is added by a propositional analysis,other than
the distinctly important fact that all learning is subsumedunder a common rubric.
Anderson and Bower’s belief that the extension of their theory to verbal learning
materialswasnecessaryto generateinterest and experimental testing of their theoretical
notions seemsto me to be wrong. The interest in Anderson and Bower’s book will
surely be primarily from cognitive psychologists, psycholinguists, linguists, and
computer scientists interested in theories of semantic memory, not verbal learning
researchers.
Also, I fail to seehow two obviously brilliant people such as John Anderson and
Gordon Bower can waste their time studying a hopelesslyuncontrolled procedure
such asfree-recall. Basicresearchtheoriesof such uncontrolled and complex processes
are a scientific absurdity. Since free-recall alsohaslittle practical significance,I consider
FRAN a wasteof time.

RECOGNITION

The goal of a recognition procedure is to map an input sentenceor someportion

thereof into someor all portions of an existing node structure (network of nodes and
associations)in semanticmemory. Recognition is an important part of the acquisition
of new knowledge in semantic memory, since new knowledge should be associated
to old knowledge. Recognition is alsoan important part of any memory usageprocess,
sincethe answeringof questionsrequires one to recognize the question as referring to
someportion of semanticmemory. HAM’s recognition processbasicallyconsistsof two
parts: the parserand the MATCH process.
HAM’s parser is a top-down, left-to-right predictive parser, which works only
becauseof the simple subsetof English sentencesgiven to it. HAM’s parser is neither
original nor powerful and has the common flaw of existing parsersin that it doesnot
take semanticknowledge into account in the parsing process.
The result of the parsing processis an input tree that has the sameform as trees
stored in semanticmemory. That is to say, HAM’s parserdoesnot produce a surface
256 WAYNE A. WICKELGREN

structure, but rather a deep structure (here used to be equivalent to a semantic

memory). This seems nice, but my intuition says that the integration of semantic
memory and the parsing aspect of sentence recognition must be much more complete
than in HAM. In HAM, parsing creates a tree structure that is initially isolated
from the rest of semantic memory and then MATCHed to existing semantic memory
structure only after completion of parsing the entire sentence.
HAM implements Anderson and Bower’s belief that syntactic knowledge should
not be represented in the same way as other knowledge in semantic memory. Maybe
this is so, and maybe it is not. My guess is that an adequate recognition routine will
operate more from the bottom-up (starting with word nodes) and will use existing
semantic memory structure from the beginning. Parsing probably proceeds in intimate
combination with mapping input words, phrases, and propositions onto existing
semantic memory word concept, phrase, and propositional nodes whenever possible,
creating new nodes when needed. I doubt that parsing, matching, and acquisition are
temporally separated at the propositional level, though they may well be somewhat
separate processes that are capable of independent study.

ACQUISITION

Anderson and Bower have little to say concerning acquisition that was not already
said in conjunction with the recognition process. This occurs for several reasons.
First, Anderson and Bower have no theory of the dynamics of parsing or acquisition.
They do have a theory of the dynamics of MATCHing, which will be discussed
briefly in the Retrieval section.
Acquisition of new nodes and links in semantic memory is accomplished in HAM
by the MATCH process which simply grafts all the unMATCHed portions of the input
tree onto the MATCHed portions following the rule of achieving the highest degree
of node economy.
However, as mentioned previously, HAM’s MATCH process is a bit overzealous
in its efforts to create higher-level node economy in that it creates compound proposi-
tional nodes. This produces the problem of independent contradiction of each proposi-
tion. Although Anderson and Bower did not acknowledge this problem, they did
note that their MATCH process can lead to “multiplication of meanings” (discussed
on pp. 243-246). As an example of this, consider that in HAM a propositional node
may join subject-l with predicate-l, later join subject-l to predicate-2, and then later
join subject-2 to predicate-2. When this happens, one also gets subject-2 joined to
predicate-l, which is not a valid inference. To solve this problem, Anderson and
Bower created another process, IDENTIFY, which inhibits the MATCH process
whenever it would produce (unwanted) multiplication of meanings. Since IDENTIFY
is a very ad hoc process which does not solve the problem of independent contra-
 SUBPROBLEMS OF SEMANTIC MEMORY 257

diction, a modification of the MATCH process to achieve structures such as shown

in Figure lb would seem to be well advised.
Avoiding multiplication of meanings raises the general principle that one good way
to evaluate any theory of semantic memory is to see, not only that it produces correct
inferences, but also that it does not produce incorrect inferences. As a consequence
of what appears to be a clear theoretical error in representation and in the MATCH
process, Anderson and Bower have generated a nice example of how this evaluation
procedure might be employed in developing theories of semantic memory.

STORAGE

HAM has four different types of memory. First, there is the sensory buffer that
contains up to seven words. Second, there is a push down store that indicates where
HAM is at in the parsing process. Third, there is a working memory that holds the
tree structures generated during parsing and for use during the MATCH and
IDENTIFY processes. Finally, there is the long-term semantic memory. This is an
excessive number of different memories for my taste. In my opinion, the reasons for this
large number derive primarily from the separation of parsing, matching, and acquisi-
tion, top-down parsing, and not enough parallel processing.
Although Anderson and Bower devote an entire chapter to the topic interference and
forgetting, it is clear that Anderson and Bower’s primary efforts were directed to
storage. Associations in HAM are “all or none,” although the ordering of associations
with identical link labels on the GET lists at each node does provide an unusual kind
of gradation. The only storage process to be found in HAM is that new associations
with the same link label at a node “bury” old associations in the recency-ordered GET
list. Retrieval at a given node involves serial search in both recognition and recall, and
Anderson and Bower assume that retrieval stops after some randomly determined
period of time. Hence, HAM has a interference mechanism that is similarity dependent
in a reasonable way.
HAM does make one unique prediction concerning storage which is that inter-
ference obtained from repeating the same word in different sentences should be obtained
only if the concept denoted by that word has the same relation to the other concepts
presented in the sentence. For example, HAM expects interference if the same concept
is used as the subject (or the agent) in two sentences, but does not expect any inter-
ference if the same concept is used as the subject in one sentence and the object in
another sentence (or the agent in one sentence and the object in the other sentence).
HAM predicts both relation-specific negative transfer and retroactive interference.
Anderson and Bower obtain the effect in transfer, but not in retroactive interference.
The prediction is obviously “up in the air. ” I was no more successful than Anderson
and Bower in figuring out why this should have happened.

480/13/3-z
258 WAYNE A. WICKELGREN

The most obvious flaw in HAM’s storage dynamics is that the ordering of the GET
lists is by recency only. However, I doubt that accounting for frequency effects would
pose insuperable difficulties for HAM. Overall, it appears that Anderson and Bower
had no desire to devote much of their theoretical effort to precise formulation of
storage processes, being content to postulate some process that was a crude first
approximation which they hoped would not greatly affect their evaluation of the parts
of the theory into which they put more effort. This is a reasonable and necessary way
to go about attacking a complex problem.

RETRIEVAL

I use retrival to refer to yes-no (recognition) and wh-(recall) question answering.

Following Anderson and Bower, this elementary retrieval will be limited to the
answering of such questions on the basis of information that was presented to a subject
in a single proposition for direct encoding in semantic memory. Answering these
questions requires no inference (combination of separately stored information).
Besides the intrinsic interest of the problem, a precise theory of elementary retrieval
is important in that it permits one to evaluate theories of the structure of semantic
memory via predictions concerning the asymptotic accuracy or dynamics of such
elementary question answering.
EssentialIy all of Anderson and Bower’s theory of retrieval is contained in the
recognition process which matches a probe tree to a portion of semantic memory.
However, in order to predict times for recognition responses (verification times)
and recall (answers to wh-questions) it was necessary to make a number of quantitative
assumptions concerning the MATCH process and the decision rule concerning
whether the retrieved structure matches sufficiently to make a “yes” response in the
case of recognition. In recall, Anderson and Bower had to assume stimulus-input and
response-output bias parameters.
Most people, myself included, feel that Anderson and Bower’s chapter 10 concerned
with recall (wh-question answering) after sentence learning was completely incon-
clusive as a means of evaluating either HAM’s structural assumptions or its assumptions
on wh-question answering, because of the manifold opportunities for compensating
errors in the estimation of the large numbers of parameters. Some of the qualitative
results concerning sentence recall are interesting, but since we already know from
other data that HAM’s model of within-propositional structure must be incorrect,
it is not clear what comfort should be taken in partially successful predictions. It seems
to me that more attention might have been paid to what Anderson and Bower refer
to as “unpredicted second recalls” and “recalls to inadequate stimuli,” since the
nature of these unpredicted recalls might indicate some of the deficiencies of HAM’s
theory of sentence structure or retrieval processes.
 SUBPROBLEMS OF SEMANTIC MEMORY 259

Recognition memory dynamics (verification times) is largely discussed in Chapter 12

(fact retrieval). The first experiment discussed in this chapter obtained the result that
verification time was slower for propositions whose constituents were found in other
propositions in semantic memory as opposed to sentences in which the constituents
were unique to that sentence. The MATCH process in HAM and in any associative
theory of semantic memory will develop multiple associations to the same concept
used in several propositions. This leads to the possibility of associative interference.
Anderson and Bower interpret this finding as evidence for a serial search of identically
labeled links from a given node. However, in view of the fact that A-C interpolated
learning produces unlearning of A-B associations (as assessed by recognition tests),
it would appear to be at least as reasonable to assume a parallel search process with
lower levels of strength for A-B associations that are followed by A-C interference.
Because of the enormous degree of speed-accuracy tradeoff found in recognition
memory retrieval (Reed, 1973, 1974) it is not possible to draw conclusions regarding
the dynamics of memory retrieval processes on the basis of reaction time data alone
without also considering accuracy level (Pachella, 1974; Wickelgren, 1974, 1975),
Speed-accuracy tradeoff in memory retrieval occurs over a substantial period of time
on the order of 500 to 1000 msec. (Reed, 1973), and the form of the speed-accuracy
tradeoff function is such that at low levels of error rate, small differences in error rate
can translate into tens or hundreds of msec. of difference in reaction times. It is precisely
at low error rate where the possibility is greatest for drawing erroneous conclusions
on the basis of reaction time alone. Furthermore, the existence of substantial speed-
accuracy tradeoff renders meaningless the quantitative prediction of reaction time as
the dependent variable, without simultaneous quantitative prediction of error rates.
Anderson and Bower’s attempt to fit a quantitative model for yes-no recognition
memory times provides yet another illustration of the irregularities one encounters
when one develops and tests quantitative theories for reaction time without considera-
tion of speed-accuracy tradeoff. In the first place, the quantitative fit of their model is
not too good. In the second place, there is substantial parameter variation. In the
third place, as Anderson and Bower move from one recognition memory reaction time
experiment to another, they shift their assumptions regarding whether the scan is
exhaustive or self terminating. Such unsatisfactory results are precisely what one
expects when a basic factor such as speed-accuracy tradeoff is being totally ignored.
Some interesting experimental results are reported in the rather curious but
interesting Chapter 11. These results are supportive of both HAM’s structure of
semantic memory and its retrieval process. However, since HAM’s theory of within-
propositional structure appears to be wrong on the basis of other evidence, these results
can only be taken as phenomena that any new theory must predict.
One interesting result was that the probability of recalling the object to the subject,
for example, was independent of the number of sentences in which the verb was
involved. So verb repetition does not produce associative interference for “irrelevant”
260 WAYNE A. WICKELGREN

constitutents of the proposition. Whether this is contradictory to a Gestalt theory

seems to me to be undecidable, since Gestalt theory is so vague and insufficiently
developed that it is not clear what predictions it makes. However, the result would
seem to be contradictory to some versions of case grammar in which the verb plays the
central role in linking subject and object. However, if one uses a different node for a
verb in every different proposition in which it occurs, as I guess Norman and Rumelhart
do, such a case theory of the structure of semantic memory would not be contradicted
by this result. Nevertheless, the result seems important.
Other interesting results were those involving crossover-cues (pp. 337-341). If a
subject learns two sentences such as “The child hit the landlord” and “The minister
praised the landlord,” HAM predicts that object recall will be slightly superior for a
sentence that mixes the subject for the first sentence with the verb for the second
(e.g., “The child praised the -,“) than for a sentence that takes the subject and
the verb from the same sentence (e.g., “The child hit the -.“). Although the
difference is not predicted to be large, it seems to me that intuition would have
suggested the opposite conclusion. In fact, the experiment showed a small difference
in the direction predicted by HAM. The result appears to mean that having two
independent cues for eliciting recall is superior to having two dependent cues (utilizing
some of the some links). This probably says more about retrieval processes involved in
wh-question answering than it does about the structure of semantic memory.
One might well question the ultimate psychological validity of distinguishing fact
retrieval for previously stored information from retrieval that requires inference.
Anderson and Bower distinguish these two subproblems, but it may well be that
precisely the same retrieval processes are involved in each case. Regardless of the
ultimate outcome of this question, it seems useful at the present time to distinguish
these two subproblems, keeping in mind the attractive possibility of eventually
developing a unified theory for both.

INFERENCE

Anderson and Bower admit that HAM is weak in inferential capacity. However,
largely in Chapter 10, Anderson and Bower list and discuss a large number of semantic
inferences that people are capable of making.

Synonyms and Definite Descriptions

Apparently, HAM is programmed to achieve the inference that “If George
Washington lived at Mt. Vernon and George Washington was the first president of
the United States then the first president of the United States lived at Mt. Vernon.”
HAM draws these inferences automatically at the time of input by virtue of having
a single representation for the concept of George Washington regardless of whether
 SUBPROBLEMS OF SEMANTIC MEMORY 261

he was referred to directly by name or by a description. The same holds for synonyms.
It is quite desirable that HAM has this inferential capacity, but it is also clear from the
results on pages 248-251 that subjects are capable of some differential encoding
for sentences with names as opposed to sentences with definite descriptions. There is
no elegant accounting for this difference in HAM, though Anderson and Bower often
make noises to the effect that differences like this could be accounted for by means of a
certain degree of auxiliary propositional encoding.
Negation
Anderson and Bower discussthe representationof negation and inferencesinvolving
negations. HAM encodespropositional negation by embeddinga negatedproposition
P asa predicate in a superordinateproposition, “It is falsethat P.”
This is okay for saying no to P in fact retrieval, but how, in general, doesone make
the inference that Q is false becauseit is contradictory to stored information. Perhaps
this should be considereda different problem.
Another nice feature of HAM is that it includes encoding of the presuppositionsof
negation. Anderson and Bower provide the example of “It wasn’t in the park that the
hippie touched the debutante.” HAM’s encoding of this is equivalent to “At some
placethe hippie touched the debutante and it is falsethat the place wasthe park.” One
presupposesthat the episode occurred and then denies that its location was in the
park. Besideits intuitive plausibility, there is somedirect psychological evidence for
the distinction between given information (the presuppositions)and the new informa-
tion in a variety of sentencetypes (Haviland and Clark, 1974).
Veri$cation vs. recognition memory. Recognition memory experiments and verifica-
tion experiments appear similar in that both involve two alternative answersthat at
first glancemight appearto be equivalent. Anderson and Bower implicitly regard them
asequivalent probably becausetheir verification experiments were essentiallyrecogni-
tion memory experiments. However, in general, verification (judging truth or falsity)
is not equivalent to recognition (judging occurrence), though there is probably a close
relation. In a verification experiment, one might present two contradictory sentences
with subjects instructed to go by the more recent sentencein determining truth. In
which case,the answer “true” and the answer “yes” (sentencehas occurred) would
have different logical determination. It is even less appropriate to identify “false”
with “no,” since “false” meansthat the probe is contradictory to stored information,
while “no” meanssomethingequivalent to “not stored in memory.” Consideration of
this issueis the sort of thing that leadsone to speculateabout the utility of the three-
valued logic where a proposition can be true “true,” “false,” or “undetermined.”
Opposites
Consider oppositesboth on binary scales,such as “in-out,” “open-closed,” “same-
different,” etc., where one term implies the negation of the other and as ends of
262 WAYNE A. WICKELGREN

continua, such as “tall-short,” “heavy-light,” “top-bottom,” etc., where the two

terms are contradictory, but the negation of one end of the continuum does not imply
the opposite end. HAM is not currently programmed to perform relevant inferences
involving opposites, but Anderson and Bower think it could be, and their discussion
is quite interesting.

Mutually exclusive concepts

Anderson and Bower refer to mutually exclusive concepts as disjoint sets. An
example inference here is that if Chicago is located in Illinois, it cannot be in Florida,
since being in Illinois implies not being in Florida. One might regard this as a simple
extension of the inferences involving ends of a continuum and perhaps an adequate
theory of mutually exclusive concepts would handle both binary and continuum
opposites and negation as well.

Set Inclusion
An example of such inferences is: “All pets are bothersome, Fido is a pet, therefore
Fido is bothersome.” As Anderson and Bower discuss, such set inclusion inferences
could be programmed into HAM, but they appear to be rather awkward to achieve. By
contrast, Quillian’s (1968, 1969) much more limited model of semantic memory
achieves set inclusion inferences quite simply and automatically. In my opinion, this
deficiency of HAM is a critical one to focus on in modifying the theory.

All Implies Some

Such inferences have the form: “If all x are P, then some x are P.” It was not clear
whether HAM was already programmed to draw this inference, but it was quite clear
that it could be. However, while the rule is trivial for this class of inferences, the
conclusion begins to dawn on one that HAM is handling a large variety of inferences
by means of a large variety of rules. It would be more elegant if a relatively small
number of rules handled all the different types of inferences people make.

Exemplar Implies Some

A seventh inference type which Anderson and Bower do not discuss until the
epitaph chapter as a deficiency of HAM is that “If xi is P, then some x are P. The
example given was “If red headed chickens are good layers, then some chickens are
good layers.” HAM does not make this inference, and clearly it should.

Disjunction (0~)

For example, “If John drove Mary home, then John or Keith drove Mary home.”
This does not seem like a very useful example of disjunctive inference, but human
beings clearly have this capability.
 SUBPROBLEMS OF SEMANTIC MEMORY 263

Expansion
These are the important classof inferencesmost cogently identified by Schank(1972)
and perhaps best described in the paper by Schank and Rieger (1973). My favorite
example of such inferencesis as follows: From “John fell overboard,” we are capable
of inferring (perhaps incorrectly but that is no matter) that “John was in a boat;
John is now in the water; John is wet; John is swimming or otherwisestruggling in the
water, etc.”
Schank handlessuch inferencesby meansof encoding at the time of the input that
expressesan extended meaning for a proposition in terms of a primitive set of atomic
concepts. I think Anderson and Bower are correct to disagreewith &hank’s approach
to this problem. Anderson and Bower claim that many of theseinferencesare not made
at the time of input, but only at the time of retrieval. They even cite someverification
latency data in support of their prediction, though I think such experiments should be
done using speed-accuracytradeoff methodology. In any event, my intuition agrees
with Anderson and Bower that many inferences are made at the time of retrieval.
However, HAM hasno mechanismfor achieving such inferences.Since human beings
make these expansions(relational and predicate inferences) easily and frequently in
understanding written or spokenlanguage,an explanation for this inferential capacity
is another primary unsolved problem.
Transitive Relations (Linear Spatial Orders)
These inferences are of the form: “A is greater than B and B is greater than C,
therefore, A is greater than C.” HAM canencodesuchlinear orderingsin an economical
way by storing the propositions “A greater than B” and “B is greater than C.” HAM
then can usea reasonablysimple, but ad hoc, inference rule for the classof transitive
relations.
One of the problems of this theory is that it makesthe production that verification
for a proposition such as, “An elephant is bigger than a termite” would be slowerthan
verification for a proposition such as, “An elephant is bigger than a tiger.” The
proposition “elephant is bigger than tiger” is presumedto be stored directly, but the
proposition “elephant is bigger than a termite” must be verified by the rule applied to
a chain of transitive relations. Anderson and Bower discusssomeevidence collected
by Potts (1972) which is basically contradictory to this prediction and attempt to
explain away Potts’ results. The paradigmsused by Potts (1972, 1973)may not yield
evidence clearly contradictory to the prediction that decisionsconcerning close pairs
are made more quickly than decisionsconcerning remore pairs in a linear ordering.
However, the results of Moyer (1973) suggestthat there is a classof linear ordering
inferenceswhich possesses the property that decisionsare faster for remote pairs than
for close pairs. Such results are incompatible either with Anderson and Bower’s
assumptionsregarding the representation of linear orderings or with their inference
and retrieval assumptions.
264 WAYNE A. WICKELGREN

GENERATION

Much of our usage of semantic memory is in generating connected speech in

conversations and soliloquies (and occasionally in lectures, dictating or writing
papers, etc.). Anderson and Bower specifically avoided developing any theory of written
or spoken sentence generation. The question is more complex than simple yes-no
or wh-questioning for at least two reasons. First, the output is more complex than a
single word or phrase as in yes-no and much wh-question answering. Problems of
mapping underlying semantic memory structure into surface grammatical structure
using syntactic rules come into play.
Second, the input conditions for such a generation of syntactically structured
utterances are not so clearly defined as they are in verification or wh-question answering.
What should we assume is the goal in speech generation ? One reasonable solution to
this problem is to imagine that the goal is to communicate fully some subset of the
semantic memory network via one or more sentences. This is probably not a totally
satisfactory answer to the input question, since one probably needs to place restrictions
on the types of subsets of semantic memory that can be the input to the generation
process. For example, it would seem unreasonable to require an adequate theory of
speech generation to be able to translate any random set of nodes and links in semantic
memory into connected discourse. It must be impossible to achieve a sentence with
grammatical surface structure for most randomly selected subsets of semantic memory.
Hence, it is necessary to specify some sort of grammar of appropriate input subsets
from semantic memory.
Then, of course, it is necessary to specify the nature of the speech generation process
that maps that subset of semantic memory into one or more grammatical surface
structures. Note that I am ignoring the supposed fact that most spoken speech is not
grammatical. For the moment, we might consider this to be a performance, rather than
a capacity, problem and ignore it. However, psychologists are property suspicious of
such solutions, and it may be that we need to know more about the real grammar of
spoken speech than the “ideal” grammar of surface structure as developed from
linguistic intuiti0n.l
The only specific suggestion that I have for solving the problem of mapping under-
lying semantic structure into surface structure is to use augmented transition networks.
For a discussion of augmented transition networks see Kaplan (1975), Stevens and
Rumelhart (1975), and the classic paper by Woods (1970). Curiously enough,
augmented transition networks appear to have been used primarily for speech recogni-
tion than for speech production, where they seem less obviously suited. Of course, it

1 Recent research by Deese (1975) indicates that most of the spoken speech of educated adults
consists of grammatical sentences, contrary to earlier intuitive assertions. Thus, performance
appears to be much closer to capacity than previously supposed.
 SUBPROBLEMS OF SEMANTIC MEMORY 265

has always been an attractive hypothesis that the same syntactic knowledge used in
speech production is also used in speech perception. Personally, I find the control
ordering characteristic of augmented transition networks to be far more natural for
speech generation than for speech recognition and reading. In any event, augmented
transition networks are quite elegant and basically consistent with an associative
memory structure.

IMAGERY

Anderson and Bower use the ideas of Winston (1970) that images are represented
not as analog pictures, but rather as sets of propositions that describe various aspects
of the image. The propositional representation of images is a very exciting idea which
currently seems open to extensive theoretical development and ultimately experimental
testing. The analog picture notion has produced some interesting experiments but has
so far defied precise theoretical statement (Pylyshyn, 1973).
Anderson and Bower specifically reject the dual encoding position that there is both
a verbal semantic memory and a spatial image memory, and instead opt for a single
unified modality of semantic memory in which the construction of images constitutes
only a greater “unpacking” of visual detail from higher order nodes in semantic
memory. This is an interesting hypothesis, but I do not see how it will wash with all
the evidence from split brain studies, lateralized brain damage, temporary immobiliza-
tion of one hemisphere, and all the other evidence indicating that verbal memory is a
separate modality from spatial image memory.
There are several different issues here. First, there is the question of whether the
representation of images is propositional or analog-pictoral. Second, even if image
memory is propositional like semantic memory, there is the question of whether the
propositional structures are equivalent in semantic and image memory. Finally, there
is the issue of whether these are two anatomically and psychologically separate
representation modalities in which there might be dual representation.
One possibility is to consider that image memory is propositional, but that the nature
of the propositional structure for the two modalities and/or the nature of the inference
processes that accompany these structures are quite different. The nature of this
difference should probably be that semantic memory has greater flexibility for repre-
sentation of various types of information, including spatial information, but that
image memory is particularly well suited for encoding, retrieving, and deriving
inferences about spatial relations.
I have a strong hypothesis concerning what the difference might be. Semantic
memory has an extensive degree of hierarchical structure for the storage of propositions
that are embedded in other propositions. Furthermore, within-propositional encoding
266 WAYNE A. WICKELGREN

probably has some hierarchical structure. In general, this sort of complex embedding
capability is quite essential to representation of human knowledge.
However, perhaps image representation has no propositional embedding and uses
a one-level symmetric relational grammar. Thus, in image memory a scene is encoded
as a set of relations with arguments, Ri(... xj . ..). where the arguments must be atomic
elements, not propositions containing relations. There is clearly a massive degree of
interconnection of various elements of the scene, since each object or part of an object
has multiple relationships to a variety of other elements. However, it appears to me
that scene description does not require embedding. I would appreciate a clear counter
example.
If no counter example is forthcoming, this is an attractive idea for the difference
between semantic memory and image memory, since a representation system that
avoided embedding would appear to facilitate information retrieval. Even if there is a
need for a certain degree of embedding for adequate scene description, it is conceivable
that such embedding is accomplished by the verbal semantic memory with the rest of
the information being stored in image memory. This begins to slip into a position
similar to that held by Anderson and Bower, but preserves the claim that there is
something basically different about image memory from semantic memory, besides the
amount of visual detail.
For the moment it seems like the propositional approach to representation of image
memory is more productive than the analog-picture approach. However, Anderson and
Bower admit that there are certain phenomena, such as Stromeyer’s (1970) eidetiker
and the Shepherd and Metzler (1971) mental rotation phenomena, that appear on the
surface to be more consistent with an analog picture approach. My guess is that we
will soon have a number of theories of image memory, with at least a few having
some analog-picture characteristics, even if the highest level image representations are
propositional.

CONCLUSION

I hope it is obvious that I regard Anderson and Bower’s book as a monumental

accomplishment. The failures of various portions of HAM are often as interesting as its
successes, and perhaps even more important. Human Associatiwe Memory is the most
valuable book I have ever read in psychology. I say this despite my guess that HAM
will rapidly be superseded by better theories of semantic memory. But no matter how
little of HAM is found in later theories, these theories will owe an enormous debt to
the intellectual work done by John Anderson and Gordon Bower. However, on a
lexical-graphemic level, I must declare that Anderson and Bower’s book is frequently
“unprincipaled.”
 SUBPROBLEMS OF SEMANTIC MEMORY 267

REFERENCES

DEESE, J. Thought into speech: statistical studies of the distribution of certain features in
spontaneous speech. Paper presented at the Psychonomic Society Meeting, November, 1975.
ESTES, W. K. An associative basis for coding and organization in memory, In, A. W. Meltin &
E. Martin (Eds.), Coding processes in human memory. Washington, D.C.: Winston, 1972.
Pp. 161-190.
FILLMORE, C. J. The case for case. In E. Bach & R. T. Harms, Uniwersuls in linguistic theory.
New York: Holt, Rinehart, & Winston, 1968. Pp. l-88.
HAVILAND, S. E., & CLARK, H. H. What’s new? Acquiring new information as a process in
comprehension. Journal of Verbal Learning and Verbal Behavior, 1974, 13, 512-521.
KAPLAN, R. M. On process models for sentence analysis. In D. A. Norman & D. E. Rumelhart,
Explorations in cognition. San Francisco: W. H. Freeman, 1975. Pp. 197-223.
MILLER, G. A. The magical number seven, plus or minus two: Some limits on our capacity for
processing information. Psychological Review, 1956, 63, 81-97.
MOYER, R. S. Comparing objects in memory: Evidence suggesting an internal psychophysics.
Perception and Psychophysics, 1973, 13, 180-l 84.
NORMAN, D. A., & RUMELHART, D. E. and the LNR Research Group. Explorations in cognition.
San Francisco: W. H. Freeman, 1975.
PACHELLA, R. G. The interpretation of reaction time in information processing research. In
B. Kantowitz (Ed.), Human information processing: tutorials in performance and cognition.
Hillsdale, N. J.: Lawrence Erlbaum, 1974. Pp. 41-82.
POTTS, G. R. Information and processing strategies used in the encoding of linear ordering.
Journal of Verbal Learning and Verbal Behavior, 1972, 11, 727-740.
POTTS, G. R. Memory for redundant information. Memory and Cognition, 1973, I, 467-570.
PYLYSHYN, 2. W. What the mind’s eye tells the mind’s brain: A critique of mental imagery.
Psychological Bulletin, 1973, 80, l-24.
QUILLIAN, M. R. Semantic memory. In M. Minsky (Ed.), Semantic information processing.
Cambridge, Mass.: M.I.T. Press, 1968.
QUILLIAN, M. R. The teachable language comprehender: A simulation program and theory of
language. Communications of the Association for Computing Machinery, 1969, 12, 459-476.
REEK, A. V. Speed-accuracy tradeoff in recognition memory. Science, 1973, 181, 574-576.
REED, A. V. The time-course of recognition in human memory. Ph.D. dissertation for the
University of Oregon, August 1974.
RUMELHART, D. E., LINDSAY, P. H., & NORMAN, D. A. A process model for long-term memory.
In E. Tulving & W. Donaldson (Eds.), Organization of memory. New York: Academic Press,
1972. Pp. 197-246.
SCHANK, R. C. Conceptual dependency: A theory of natural-language understanding. Cognitive
Psychology, 1972,3,552-631.
SCHANK, R. C. Identification of conceptualizations underlying natural language. In R. C. Schank
& K. M. Colby, Computer models of thought and language. San Francisco: W. H. Freeman,
1973. Pp. 187-247.
SWANK, R. C., & RIEGER, C. J. Inference and the computer understanding of natural language.
Stanford Artificial Intelligence Memo AIM-197, May, 1973.
SHEPHERD, R. N., & METZLER, J. Mental rotation of three-dimensional objects. Science, 1971,
171, 701-703.
STEVENS, A. L.. & RUMELHART, D. E. Errors in reading: An analysis using an augmented transi-
tion network model of grammar. In D. A. Norman and D. E. Rumelhart, Explorations in
cognition. San Francisco: W. H. Freeman, 1975. Pp. 224-263.
268 WAYNE A. WICKELGREN

STROMEYER, C. F. Eidetikers. Psychology Today, 1970, 4, (November) 76-81.

WICKELGREN, W. A. Size of rehearsal group and short-term memory. Journal of Experimental
Psychology, 1964, 68, 431-419.
WICKELGREN, W. A. Short-term memory for repeated and non-repeated items. Quarterly Journal
of Experimental Psychology. 1965, 17, 14-25.
WICKELGREN, W. A. Rehearsal grouping and hierarchical organization of serial position cues
in short-term memory. Quarterly Journal of Experimental Psychology, 1967, 19, 97-102.
WICKELGREN, W. A. Learned specification of concept neurons. Bulletin of Mathematical Bio-
physics, 1969, 31, 123-142.
WICKELGREN, W. A. Coding, retrieval, and dynamics of multitrace associative memory. In
L. W. Gregg (Ed.), Cognition in learning and memory. New York: Wiley, 1972. Pp. 19-50.
WICKELGREN, W. A. Speed-accuracy tradeoff and information processing dynamics. Paper
presented at The Psychonomic Society Meeting, November, 1974.
WICKELGREN, W. A. Dynamics of retrieval. In D. Deutsch and J. A. Deutsch (Eds.), Short-term
memory. New York: Academic Press, 1975. Pp. 233-255.
WICKELGREN, W. A., & WHITMAN, P. T. Visual very-short-term memory is non-associative.
Journal of Experimental Psychology, 1970, 84, 277-281.
WINOGRAD, T. Understanding natural language. New York: Academic Press, 1972.

RECEIVED: November 1 I, 1974