Only Connections: A Critique of Semantic Networks

Psychological Bulletin1984, Vol. 96, No. 2, 292-315

Copyright 1984 by theAmerican Psychological Association, Inc.

Only Connections: A Critique of Semantic Networks

P. N. Johnson-LairdMedical Research CouncilApplied Psychology Unit

Cambridge, England

D. J. HerrmannHamilton College

R. ChaffinTrenton State College

This article examines theories that assume that semantic networks account forthe mental representation of meaning. It assesses their similarities and divergencies,and argues that as a class of theories they remain too powerful to be refuted byempirical evidence. The theories are also confronted by a number of problematicalsemantic phenomena that arise because networks deal with the connectionsbetween concepts rather than with their connections to the world. The solutionto these problems could be embodied in a new network system, but such asystem would differ in both structure and function from current network theories.

In the past 15 years, a number of memory-and linguistic-processing theories have beenadvanced with a common approach to se-mantics: They have assumed that meaning isrepresented by a network of labelled associ-ations. These so-called semantic network the-ories have stimulated considerable research,both experimental studies of categorizationand computer implementations of networks.However, no systematic investigation has beenmade of semantic networks as a class ofpsychological theories. The aim of this articleis to provide such an examination in orderto assess the strengths and weaknesses of thisapproach to the psychology of meaning.

Semantic networks are used to model per-formance in a variety of tasks. A reviewcould examine them from several angles (e.g.,how they explain verbal learning, or howthey account for the comprehension of prose).We choose to concentrate on linguistic mean-ing because network theories have made animportant contribution to elucidating the se-

The authors are grateful to Doug Hintzman, RichardYoung, and two anonymous reviewers for some stimulatingcomments on an earlier draft of this article. The secondauthor thanks the MRC Applied Psychology Unit and itsdirector, Alan Baddeley, for providing a congenial atmo-sphere for research while he was a visitor there in 1982-1983.

Requests for reprints should be sent to P. N. Johnson-Laird, MRC Applied Psychology Unit, 15 Chaucer Road,Cambridge CB2 2EF, England.

mantic representation of words and sentences.We are concerned primarily with the empir-ical content of the theories rather than withthe formalism of networks, although inevi-tably we cannot consider one without theother.

In this article, we first outline the maingoals to be achieved by any cognitive theoryof meaning. Second, we review the mainnetwork theories that are intended as psycho-logical theories. Third, we consider what it isthat such theories have in common and whatpotential constraints may be placed on theclass of network theories as a whole. Fourth,we outline a variety of semantic phenomenathat network theories have difficulty in ac-commodating. Fifth, we consider what lies atthe root of these phenomena and advance apotential explanation of them. Finally, wedraw some brief conclusions about the re-sulting status of semantic networks in relationto the goals outlined in the first section.

Goals of a Psychological Theory of Meaning

Natural language enables human beings tocommunicate ideas. For example, if a speakerasserts in a conversation about ChristopherColumbus, "The captain of the Santa Mariathought the earth was round," listeners areable to recover the significance of the remark.This simple example of communication stilldefies analysis and provides us with a usefultest case. It illustrates most of the major goals

292

SEMANTIC NETWORKS 293

for a psychologically plausible theory ofmeaning (see Bierwisch, 1970; Lyons, 1977;Miller & Johnson-Laird, 1976, p. 706).

In our view, there are four principal goals.The first is to specify the form of the mentalrepresentation of meaning. Such a represen-tation is in the mind of the speaker andmapped into words by the speech productionsystem. The listener decodes these words intoa mental representation which, if all goeswell, captures what the speaker had in mind.It is important to note the distinction, drawnby Frege (1892), between sense and reference.The reference of an expression is what itstands for in the world; the sense of anexpression is that part of its meaning thatconcerns the way the expression connectswith its reference. Hence, the reference of theexpression "the captain of the Santa Maria"is Christopher Columbus, but other expres-sions with different senses, such as "the dis-coverer of America" or "the first person tosail the Atlantic," can have the same reference.Frege's informal notions of sense and refer-ence have been replaced in formal semanticsby the concepts of intension and extension,and we adopt this terminology. Thus, theintension (or sense) of a word such as captaindetermines the set of all its possible extensions(or referents), and its extension in a particularproposition, such as the one expressed by ourexample, is a particular individual. We arenot concerned with the logical analysis ofintensions and extensions, but it is crucial tobear in mind two distinct sorts of relations:the relation between the sense of one word(or expression) and another, which we call anintensional relation, and the relation betweena word (or expression) and its referent, whichwe call an extensional relation.

The second goal of semantic theory is toexplain intensional phenomena. It should ac-count for the intensional relations of wordsand expressions, such as synonymy, anto-nymy, and inclusion (e.g., captain has "hu-man" as a superordinate; the example sen-tence implies: "The captain of the SantaMaria did not think that the earth was flat").It should account for the semantic propertiesof expressions such as ambiguity, anomaly,analyticity (i.e., truth in virtue of meaning),and self-contradiction (i.e., falsity in virtueof meaning).

The third goal is to explain extensionalphenomena. The theory should account forthe extensional relations between words andthe world as human beings conceive it. Itshould elucidate the way in which speakersand listeners relate expressions to their exten-sions. It should also establish the way inwhich semantic representations capture thetruth conditions of assertions and the sensesof questions and commands. A major puzzleis how a semantic representation captures thefact that assertions are true with respect toan infinite number of different situations.Even the simple remark about Columbus istrue of many different possibilities: He mighthave thought that the world was an ellipsoid,cylindrical, pear shaped, and so on. A picturemay be worth a thousand words, but a prop-osition is worth an infinity of pictures.

The fourth goal is to explain the inferencesthat people draw in virtue of the meaningsof words. That is, the form of semanticrepresentations should dovetail with whatevermachinery is used to make verbal inferences.

There are undoubtedly other goals for se-mantic analyses, but these four are perhapsthe most crucial for a psychological theory.They all concern what is computed in varioussemantic tasks, and obviously the theory mustalso account for how the tasks are carriedout (i.e., for the various mental processesunderlying them). The theory should specifyan effective procedure (or algorithm) by whichthe mind constructs semantic representationsduring comprehension and an effective pro-cedure by which semantic representations aremapped into words during speaking. Theseprocedures presumably depend on a grammar,a lexicon, knowledge of the world, and theability to make inferences. The theory shouldalso specify procedures for evaluating inten-sional properties and relations, becausespeakers are able to judge whether an expres-sion is ambiguous, anomalous, or meaningful(e.g., Steinberg, 1970). They are also able tomake judgments of synonymy and antonymy(e.g., Herrmann, Chaffin, Conti, Peters, &Robbins, 1979). Finally, the theory shouldspecify procedures for representing and eval-uating extensions and truth values, becausespeakers can determine what an expressionrefers to and ascertain whether assertions aretrue (Clark & Clark, 1977).

294 P. JOHNSON-LAIRD, D. HERRMANN, AND R. CHAFFIN

In short, a psychological theory of meaningshould explain how meaning is mentally rep-resented, how expressions are intensionallyrelated one to another, how they are relatedto the world as the mind conceives it, andhow their semantic representations enter intoinferences. In due course, we examine howwell semantic networks meet these explana-tory goals, but first we outline this class oftheories.

Semantic Network Theories

Early Semantic Networks

Semantic networks are associative theoriesframed for computers. Associationists fromAristotle to the present day have assumedthat there can be an association from oneword to another. The set of associations to agiven word will contain many members thatwill differ in strength, and many of the wordsin the set will themselves be associativelyrelated. It is therefore natural to think ofwords as associated by a network of undiffer-entiated links varying in strength (see, e.g.,Deese, 1962, 1965; Kiss, 1967; Walpole, 1941;Warren, 1921). Although such a networkmay be explanatorily useful for studies ofword association, it is obviously a poor in-strument for semantics because a mere asso-ciative link from one word to another tellsone nothing about the intensional relationbetween the words. For example, black isstrongly associated with both white and night,but the intensional relation between blackand white is very different from the intensionalrelation between black and night.

The decisive step in converting an associa-tive network into a potential semantic ma-chine was the introduction of labels on thelinks between words. The importance of rep-resenting the relation between entities wasrecognized by Selz in 1913 (cited by Hum-phrey, 1951), but Selz lacked an appropriatelanguage in which to make his theory whollyexplicit. With the advent of programminglanguages, and particularly list-processinglanguages such as IPL and LISP, the way wasclear for the invention of semantic networks.An expression in LISP is either a list or anatom, and an atom is either a number or asymbol, such as poodle. LISP allows one to

set up a whole series of property names andtheir specific values associated with a symbolicatom such as poodle:

Superordinate: dog

Size: small

Hair: curly

In describing such information, it is naturalto represent it in a graphical form:

Superordinate SizeDog < Poodle —> Small

Hair J.Curly

Early semantic networks (Quillian, 1968; Ra-phael, 1968) were probably inspired in partby this feature of IPL and LISP.

Quillian's (1968) theory anticipated mostof the features of subsequent semantic net-works. He assumed that memory for meaningis no different from memory for perceptualor other nonlinguistic information, and hepostulated a semantic network as a modelfor lexical memory. The network is composedof links between two sorts of nodes: typenodes, which represent concepts, and tokennodes, which represent instances of conceptsby virtue of the links to their respective typenodes. The meaning of a word is defined byan initial configuration of token nodes at-tached to the type node representing theword, and each of the token nodes is linkedto its respective type node. Figure 1 presentsa simplified example of such a network,which represents one sense of the verb "toplant." As the figure illustrates, the theoryproposes five sorts of links: (a) From a tokento its type node (as shown by the dotted linesleading from tokens out to their respectivetypes, which are omitted from the figure); (b)from a type to a Superordinate token (e.g.,the link from plant to put); (c) from a tokento another token representing a modifyingproperty (e.g., the link from put to for torepresent the fact that seeds are put in theground for them to grow); (d) from a tokento a conjunction or disjunction of tokens(e.g., from object to seed, plant, or thing); (e)from a token to two other tokens so as to actas a label on a link between them (e.g., objectand earth are related by in because the object


,-PUT

VSUBJECT

PERSON...

\

Figure ]. A semantic network representation of one sense of the verb "to plant." (From "SemanticMemory" by M. R. Quillian, in Semantic Information Processing [p, 236] edited by M. L. Minsky, 1968,Cambridge, MA: MIT Press. Copyright 1968 by MIT Press, Reprinted by permission.)

is put in the earth). The amount of infor-mation in a network is potentially so vastthat Quillian assumed that facts are storedexplicitly only if they cannot be generatedfrom the network. Hence, general informationneed be represented only at a superordinatelevel without being attached to all the sub-ordinate nodes to which it applies. For ex-ample, a poodle is a dog and a poodle is ananimal, but because all dogs are animals, itis parsimonious to use superordinate links inthe network:

Poodle —» Dog —* Animal

There is no need to represent the fact that apoodle is an animal, because this informationcan be inferred from the network by tracingthrough the pathway from poodle to animal.

Quillian introduced various "tags" thatcould be attached to nodes. They include anumber representing the degree to which aproperty represented by one node applies toan entity represented by another node, anumber representing the "criteriality" of atoken in denning a type, and special symbolsthat indicate the role of other words in denn-ing a concept. In this last case, Quillian usedsymbols to show that one word could be thesubject or the direct object of the currentword, or that a word could be modified by

the current word. Thus, as Figure 1one sense of plant calls for a person,! assubject, putting seeds, plants, or other things,as object, into the earth. Quillian suggestedthat this machinery could be used to sehctthe appropriate meaning of the verb, whichis otherwise ambiguous, from the linguisticcontext in which it occurs. He also suggestedthat the labels of Fillmore's (1968) C»segrammar might be more appropriate thanthese standard grammatical terms.

Such a theory about representation be-comes a full-fledged theory of performanceonly when processes for using the represen-tation—for setting it up, interrogating it, anddrawing inferences from it—are specified.Indeed, as Quillian suggested, certain aspectsof meaning may be represented by suchprocesses rather than directly in the network.A natural assumption, however, is that whenpeople evaluate the relation between two con-cepts, they do so by searching through thenetwork for a path between them. Quill: .anwrote a computer program that operates inthis way. The program simulates parallel pro-cessing by making a breadth-first search alongthe links radiating out from the two typenodes activated by the input words. At eachnode, the program leaves tags specifying boththe immediately preceding node in the search


and the original starting node. Thus, thesearch moves outward in a constantly ex-panding spread of activation. An intersectionoccurs when the search from one startingpoint encounters a node with the tag fromthe other starting point. The program thenevaluates the path between the starting nodesin order to establish the nature of the inten-sional relation between them.

Quillian's model aroused interest amongpsychologists because it could be used topredict how long it takes to verify differentsorts of statement. Collins and Quillian (1969,1972) assumed that the greater the length ofthe path that has to be traversed to establishan intersection, the longer the task shouldtake, and they reported some experimentalevidence corroborating this assumption. Theyfound that subjects take longer, for example,to evaluate the assertion that "a poodle is ananimal" than to evaluate the assertion that"a poodle is a dog." However, later researchhas revealed a more complicated picture. Inparticular, some results run counter to theproposal that properties are always stored atthe most economical superordinate node(Conrad, 1972). Other results suggest that therelative sizes of sets may be more criticalthan path length (Landauer & Freedman,1968). Moreover, some instances of a conceptare more prototypical than others, and asser-tions about such typical instances (e.g., "Arobin is a bird") are verified faster thanassertions about atypical instances (e.g., "Apenguin is a bird"; Rips, Shoben, & Smith,1973; Rosch, 1973; Smith, Shoben, & Rips,1974). In short, cognitive economy is notinvariable, and path length alone does notalways determine verification time.

The semantic network as we have describedit so far constitutes a theory about the orga-nization and processing of the mental lexicon:Representations of words are stored in anetwork, and the semantic relations betweenwords are represented by labelled links be-tween the items in the network. Semanticallyrelated words have shorter paths betweenthem, and it is supposed to take less time totraverse these paths than it does to traversethose between more remotely related words.Analytic assertions, which are true in virtueof the meanings of words (e.g., "A poodle isan animal"), and self-contradictory assertions,

which are false in virtue of the meanings ofwords (e.g., "A poodle is a car"), are evaluatedby checking the relation that is representedin the network.

The next major step in the evolution ofnetwork theory was to show how the meaningof any sentence could be represented as asemantic network. A number of differenttheories were proposed to do this job, butthey have in common the following basicidea of Quillian's. If a sentence uses a verbto establish a relation between the entitiesdenoted by noun phrases, then the semanticrepresentation of the sentence can take theform of a small-scale semantic network thatcaptures the relation (expressed by the verb)between the entities (denoted by the nounphrases). The full semantic representationsof the words in the sentence are, of course,encoded separately in the main network cor-responding to the mental lexicon. The inter-pretative process setting up the representationof the meaning of a sentence must contain amechanism, such as the one previously de-scribed, for selecting the appropriate sense ofan ambiguous word as a function of thecontext in which it occurs.

Network theories differ principally in theirconfigurations of links and in the types oflabel that they use. These differences to someextent echo those in the linguistic theoriescurrent when the networks were formulated.We now examine the major psychologicalnetwork theories in more detail, concentratingon the differences between them and on theircontrasts with Quillian's seminal formulation.

Human Associative Memory(HAM) Network

Human Associative Memory (HAM) is acomputer model of long-term memory devel-oped by J. R. Anderson and Bower (1973).Like Quillian, these authors assumed thatboth linguistic and perceptual information isstored in the form of abstract prepositionalrepresentations (i.e., structures in a network).This assumption was motivated by two phe-nomena: (a) that people primarily rememberthe gist of sentences, not their details verba-tim; and (b) that subjects can make infer-ences—it was presumed that they were basedon rules of inference that operate on propo-


Context Fact

Location Predicate

isa

Word

PARK PAST HIPPIE

Object

isa

Word

TOUCH DEBUTANTE

Figure 2. The HAM representation of "In a park a hippie touched a debutante." (ISA denotes setmembership. From Human Associative Memory [p. 67] by J. R. Anderson and G. H. Bower, 1973,Washington, DC: Winston. Copyright 1973 by Winston. Reprinted by permission.)

sitional representations. J. R, Anderson andBower devised a simple parser that maps alimited subset of English into network rep-resentations. Figure 2 presents their networkfor the sentence, "In a park a hippie toucheda debutante." Each word has a correspondingconcept represented by a circle. All the se-mantic information is represented by theconfiguration of the links and the labels onthem; the nodes themselves have no semanticlabels. The sentence is represented in twoparts: the context ("in a park") and the fact("a hippie touched a debutante"). The ISAlabel denotes set membership, and the re-maining labels are reminiscent of those to befound in the standard theory of transforma-tional grammar: subject, predicate, object.As in Quillian's theory, the concept nodesare taken to exist already in memory, andthe representation of the sentence accordinglydepends on them.

In addition to the features illustrated inFigure 2, HAM also contains machinery forcoping with some aspects of quantifiers (e.g.,it has symbols denoting universal quantifica-tion and implication). The theory also as-sumes that a process of pattern matchingsearches memory for the existence of a givenproposition. J. R. Anderson, however, con-

cluded that HAM is incomplete, and hecreated a new network theory, which wedescribe shortly.

Lindsay, Norman, and Rumelhart(LNR) Model

Over a number of years, Lindsay, Norman,and Rumelhart developed a theory of long-term memory and comprehension (see, e.g.,Rumelhart, Lindsay, & Norman, 1972; Nor-man, Rumelhart, and the LNR ResearchGroup, 1975; Lindsay & Norman, 1977).Their system again decomposes into a se-mantic network, processes that operate on it,and a parser that maps sentences into anetwork representation. The semantic net-work comprises labelled nodes (the only itemsthat are directly addressable in the network)and labelled links between them. There arefour sorts of nodes in an LNR network:concept, primary, secondary, and event nodes.A concept node represents a concept, and themain links used in defining it are: ISA (forset inclusion), IS (for attributing properties),and HAS (for attributing proper parts and soon). Some concepts do not have any simplelinguistic correlates, but a primary node doesrelate directly to a word in the language, and


OBJECT ISA

< > ^DEBUTANTE

TOUCH

ACT

ISA ACTOR I

HIPPIE"* < >"* (*TOUCH

TIME// \LOCATION/ \

PAST < >

I ISA

PARK

Figure 3. The LNR representation of "In a park a hippie touched a debutante." (ISA denotes setmembership.)

may have attached to it a definition of theword (in the network format). A secondarynode represents a specific use of a primarynode (e.g., in representing the meaning of aspecific sentence). Thus, the distinction be-tween a primary and a secondary node isakin to Quillian's distinction between a typenode and a token node. Events have a specialstatus in the network and are represented byevent nodes that are linked to the actions,actors, and objects involved in the event. Aspecific event is encoded by a secondary nodewith a link, labelled ACT, to the correspond-ing primary node. Thus, concept and eventnodes are distinguishable by whether an ISAor an ACT link is attached. Figure 3 showsan LNR network representation for the sen-tence, "In a park a hippie touched a debu-tante." As the figure shows, the labels onlinks from nodes are based on the cases (e.g.,agent, object, recipient) of Fillmore's (1968)Case grammar. In addition, there are a num-ber of operations that increase the power ofthe system: operations for forming variousconnections between propositions (e.g., con-junction), operations for introducing quanti-fication, and operations that generate newrelations from old ones (e.g., to capture ad-verbs such as slowly).

Like Quillian, the LNR group assumesthat in the lexical network general propertiesare stored at just one node (e.g., the fact thatcanaries have wings is recoverable from thefact that a canary is a bird and birds havewings). However, this generic information istreated as a set of values that can be assumedby default (e.g., typically, birds fly) ratherthan as necessarily true. Hence, the system

can absorb exceptions (e.g., penguins do notfly) without risk of inconsistency.

In their later work, the LNR group arguedthat a representation such as Figure 3 issuperficial, and that a deeper level of repre-sentation is required in order to capture thefull meanings of sentences. These deeper rep-resentations decompose the superficial onesinto their appropriate semantic primitives.Figure 4 presents such a primitive meaningstructure, which is obtained by replacingitems in the superficial network by theirnetwork definitions.

Glass and Holyoak's "Marker" Theory

Glass and Holyoak (1974/1975) proposeda network theory of the lexicon, which is inessence a revision of the original Quillianmodel designed to accommodate some ex-perimental results. They suggested that se-mantic markers, similar to those postulatedby Katz (1972), form a structure that under-lies the intensional relations between words.Figure 5 presents an example of such anassociative structure. Each noun is mentallyassociated with a defining marker, which issupposed to represent an abstract conceptroughly equivalent to possessing the essentialproperties of X, where X is the noun inquestion. (How an uninterpreted symbol canserve this semantic function is not explained,and merely connecting it to other uninter-preted symbols would not seem to suffice.)There are other markers, such as pet-canary,that are not directly associated with a word.The links between markers represent relationsbetween concepts. For example, the links


FROM

< > 'POSSESS

ISA I SUBJECT

JOHN

.POSSESS;SUBJECT? ^OBJECT

ISA I ISA

MARY FIDO

Figure 4. The LNR primitive meaning representation of"John gave Fido to Mary," (ISA denotes set membership.From "A Language Comprehension System" by D. E.Rumelhart and J. A. Levin, in Explorations in Cognition[p. 193] edited by D. A. Norman, D. E. Rumelhart, andthe LNR Research Group, 1975, San Francisco: Freeman.Copyright 1975 by Freeman. Reprinted by permission.)

from avian to feathered and animate capturethe fact that avian stands for the set: {(ani-mate), (feathered)}. The links accordinglydenote set inclusion. The chief innovation isthat there are two types of intersection oflinks: consistent-sod inconsistent intersections.(The authors use the terms contradictory andnoncontradictory, which we have eschewedbecause contradictory has the unfortunateconnotation that invariably one link must betrue and the other false, whereas in fact bothlinks can be false of a given entity.) Somethingthat is animate can be avian or canine, butit cannot be both, and so these two links areinconsistent with one another. Something thatis animate can, however, be both avian and apet, and so these two links are consistentwith one another. Links that have the same

Greek symbol in Figure 5 are accordinglyinconsistent with one another, whereas linksthat have different Greek symbols are consis-tent with one another. Because the linksrepresent the transitive relation of set inclu-sion, consistency or inconsistency can ariseat a relatively remote point in the network(e.g., a canary cannot be a collie, becausethere is an inconsistent intersection at theanimate node). This scheme is a direct reflec-tion of Katz's (1972) treatment of antonymy.Inconsistencies can also arise in quite a dif-ferent way. An assertion such as "All birdsare canaries" does not yield inconsistent linksbut one can judge it to be false by thinkingof another species of birds, such as robins,that are in an inconsistent relation with ca-naries. In other words, a putative generaliza-tion can be falsified by a counterexample.

Glass and Holyoak (1974/1975) allowedthat a network need not be strictly hierarchi-cal (as did Quillian and others). There maybe, for example, a direct link from canary toanimate. As they recognized, however, theintroduction of this sort of possibility releasesthe network from any formal constraints onits configurations with the ensuing dangerthat it becomes empirically vacuous. Theytherefore tried to use empirical results toplace constraints on the theory. In particular,they used a sentence-completion task to pro-vide evidence about the order in which linksare searched. Subjects had to complete suchfragmentary sentences as "All canaries are. . ." either so as to produce a true assertionor else so as to produce a false one, and therelative frequency of the different responses

PET. -ANIMATE

pet Xanimal

FEATHERED-* — AVIAN^bird

7\~k TAHADY DntllM* UAINAKT KUDIPl

t tcanary robin

CANINE<=dog

COLLIEt

coltie

PET-CANARY

Figure 5. An example of an associative network for semantic markers. (Words are represented in lowercaseletters, and markers in capital letters. Links labelled with the same Greek letter are inconsistent with oneanother, whereas links labeled with different Greek letters are consistent with one another. From"Alternative Conceptions of Semantic Memory" by A. L. Glass and K. J. Holyoak, 1974/1975, Cognition,3, p. 319. Copyright 1974/1975 by the Associated Scientific Publishers. Reprinted by permission.)


was treated as an index of search order. Thisorder predicts the latencies of evaluating an-alytic assertions, because subjects should re-spond faster to an assertion of the form "AllAs are Bs" than to one of the form "All Asare Cs" if the A —> B link is searched priorto the A —> C link. (Nothing hinges on thenotion that one link is literally searchedbefore another: In a parallel search, one linkcould be traversed faster than another.) Thetrue completions do indeed predict the laten-cies of evaluating true assertions. The troublewith this empirical constraint, however, isthat most theories, whether or not they arebased on networks, would predict the samecorrelation: It is merely Marbe's law as appliedto semantic judgments (see Johnson-Laird,1974). What is more striking are the datafrom the false sentences. Those that arisefrom an inconsistency of links (e.g., "Allbirds are dogs") were evaluated with a latencythat was predicted by the frequency withwhich the predicate occurred in the falsecompletions. Such pairs as bird and dog arejudged as being closer in meaning than pairsthat are produced relatively infrequently andevaluated relatively slowly. The results seemto run counter to the hypothesis that falseassertions containing subject and predicateterms that are judged to be semanticallyrelated take longer to evaluate, but the exper-iment has been criticized on methodologicalgrounds (see McCloskey & Glucksberg, 1979;Shoben, Wescourt, & Smith, 1978). Falsesentences that arise from the existence ofcounterexamples rather than from inconsis-tent links (e.g., "All birds are canaries")require the subject to discover a subset of thesubject (such as "robins") that is inconsistentwith the predicate. The latency of the evalu-ation should therefore depend not on thefrequency with which the sentence itself isproduced as a false completion but on thefrequency with which the counterexamplecategory occurs in true completions of "Somebirds are. . . ." This prediction was alsoborne out by the results.

Glass and Holyoak's theory is deliberatelyrestricted to relations of set inclusion, andthe strong point of their model is its detailedpredictions of the times it takes to reject thetwo kinds of false set-inclusion statements.Other semantic models, together with other

interpretations of the data, are restricted tothe single prediction that the more related inmeaning the subject and predicate of a falsestatement are, the longer it will take to rejectthat statement (McCloskey & Glucksberg,1979; Smith et al., 1974). The principles ofthe Glass and Holyoak model can be readilyincorporated within the more general Quil-lian-based network theory, and, in part, thenext theory that we consider was designedfor this purpose.

Collins and Loftus Theory

Collins and Loftus (1975) pointed out anumber of common misconceptions aboutQuillian's original theory, and proposed anumber of extensions to account for variousempirical phenomena. Some of these as-sumptions concern the way in which activa-tion spreads through the network; others con-cern the organization of the network itself.Collins and Loftus drew a clear distinctionbetween the lexical network, which is mainlyorganized on the basis of phonemic similarity,and the conceptual network, which is orga-nized on the basis of semantic similarity.Semantic similarity depends on the numberof properties that concepts share, and thusthe number of links between them. It isdistinct from semantic distance, which is theshortest distance between the relevant nodes:Two nodes may be close, yet not highlyrelated (e.g., cherries and fire engine arerelatively close because they adjoin red, butthey have np other links in common). Oneof the major consequences of the new theoryis that if, for example, vehicle is primed, thenall the different types of vehicles will beprimed and will prime each other, whereas ifred is primed, then fire engine and cherrieswill be primed, but there will be much lessmutual priming because they have no otherlinks in common.

The authors also outlined a revised versionof the decision process postulated by Collinsand Quillian (1972). To decide whether twoconcepts match, sufficient evidence must ac-cumulate to exceed either a positive or anegative threshold. Evidence accumulatesfrom different pathways, with pieces of posi-tive and negative evidence cancelling eachother out in an essentially Bayesian way.


Predicate Predicate

Relation/ \Argument

Subject

Relation

AT-TIME PAST HIPPIE TOUCH

Argument

Argument

DEBUTANTE AT-PLACE PARK

Figure 6. The ACT representation of "In a park a hippie touched a debutante." (Following J. R. Anderson,the diagram has omitted the nodes representing words. From Language, Memory, and Thought [p. 166]by J. R. Anderson, 1976, Hillsdale, NJ: Erlbaum. Copyright 1976 by Erlbaum. Reprinted by permission.)

Positive evidence consists of paths that estab-lish that one concept is a superordinate ofthe other, or that they share a commonproperty (depending on how criterial thatproperty is with respect to each concept), orthat one concept has a property of an instanceof the other (the Wittgenstein strategy, becauseit occurs with concepts that depend on familyresemblances among their instances ratherthan on a set of common elements, e.g.,different sorts of games). Negative evidenceconsists of paths that establish that one con-cept is not a superordinate of the other, orthat they have properties that mismatch (de-pending again on criteriality), or that oneconcept lacks the properties of instances ofthe other. In addition, following Holyoak andGlass (1975), negative evidence is providedby establishing that the two concepts aremutually inconsistent subordinates of thesame superordinate, or by finding a counter-example to the putative relation. AlthoughCollins and Loftus (197S) postulated onlythese kinds of evidence for categorizationtasks, they do allow that there may be otherkinds, particularly for answering complicatedquestions.

One consequence of these various revisionsis that the Collins and Loftus model hasmuch the same strengths and weaknessesalready attributed to the Glass and Holyoakmodel. Because the network is no longerstrictly hierarchical, there are no strong formalconstraints on its configurations. In contrastthe model can accommodate the experimentalresults that had been interpreted as contrary

to network theory. In particular, the theoryis now compatible with the existence of pro-totypes. As we have already pointed out, pathlength within the network does not alwaysdetermine verification time. Rather, set-in-clusion statements about typical instances ofa category (e.g., "A robin is a bird") areverified faster than statements about atypicalinstances (e.g., "A penguin is a bird"; Ripset al., 1973; Rosch, 1973; Smith et al., 1974).The phenomenon is now explicable on thegrounds that an atypical instance elicits neg-ative evidence to a greater "degree than doesa typical instance.

ACT

J. R. Anderson (1976) developed ACT asa more advanced theory of representationthan its precursor HAM. The theory assumesthat only a portion of the network is activeat any moment, with activation spreadingfrom one node to another,'although there isa general dampening of activation to preventit from getting out of control. Links in thenetwork can vary in their strength, and thespread of activation depends on the currentstrengths of the various links. The machineryfor manipulating the network consists of aproduction system in which several produc-tions can be applied at the same time. Atypical representation of a sentence in theACT system is shown in Figure 6. As thefigure illustrates, sentences are no longer di-vided into facts and contexts; the notation isnow more uniform and closely related to the


standard theory of transformational grammar(Chomsky, 1965; cf. J. R. Anderson, 1976, p.158). Anderson rejected the use of Case-grammar labels on the grounds that they lackcoherent semantics. Unlike the LNR system,ACT does not decompose the meanings ofwords into a network of primitives, but ratheranalyzes their semantics by using productions.The theory therefore resembles the proposalsof both Kintsch (1974) and Fodor, Fodor,and Garrett (1975) in that inferences arebased on rules rather than on decompositioninto a network of semantic primitives. Finally,ACT contains more powerful machinery forcoping with quantification, although much ofit is ad hoc (e.g., the sentence "All philoso-phers have read all of Plato's books" cannotbe directly represented in the network formatalthough it can be encoded as a production).The expressive power of ACT accordinglydepends on both the network and the pro-duction system.

What is a Semantic Network?

Assumptions Underlying Network Theory

There is no doubt that network theorieshave made a significant contribution to thestudy of semantics. They have inspired aconsiderable amount of experimentation,which has revealed new empirical phenom-ena, and they have been modeled in manycomputer programs, including systems de-vised purely for the purposes of artificialintelligence (see, e.g., Bobrow & Collins, 1975;Findler, 1979). There is such a variety ofnetwork theories, however, that a skepticmight suspect they have nothing in commonapart from the name. In fact, there appearto be four main assumptions that they sharewhich we now try to make explicit.

First, network theories are designed pri-marily to elucidate intensional relations, inparticular, the relations between the meaningsof words. They embody no general principlesconcerning extensions.

Second, and a corollary of the previouspoint, semantic networks are constructed onthe assumption that the representation andevaluation of intensional relations can beconsidered independently from extensionalrelations (e.g., dogs can be represented as a

subset of animals without worrying abouthow to characterize the extensions of the twoterms).

Third, comprehensive network theories arebased on a formalism containing three com-ponents: a parser, a semantic memory con-sisting of a network of links between nodes,and a set of interpretative processes thatoperate on the network. Intensional relationsbetween words are represented within thesemantic network by labelled links betweenthe nodes. The parser uses this informationto construct network representations of sen-tences. The interpretative processes carry outsuch tasks as updating the semantic network,making inferences from sentences, andsearching the semantic network to establishthe intensional properties of expressions andintensional relations between them. Theseaspects of intensions are a direct consequenceof what is encoded in the network system. Inparticular, the analyticity of a sentence followsdirectly from information in the network(e.g., the truth of an assertion such as "Ca-naries are birds").

Fourth, there is a general, although notabsolute, commitment to parsimony. If infor-mation about the meaning of a word can beinferred by traversing links, then it is notredundantly specified in the network. Hence,general facts can be stored at the level of asuperset rather than for each subset to whichthey apply.

These four assumptions are all that we candiscern as common to network theories, but,as we now show, some of the apparent diver-sity is superficial.

Differences Between Network Theories

The best way to distinguish between thesuperficial and the more profound differencesamong the theories is to imagine that atheorist is setting up a new network theory.Certain questions will inevitably arise aboutits organization and contents. Some of thesequestions are impossible to answer fromwithin the framework of network theory.Others are less problematical, and their an-swers will help to place some further con-straints on the set of possible network theories.There are many decisions that the theoristdeveloping a new system has to make; we


Semantic memory(JM PLICATIONS) (OWN INGS) SCARS'

ANTECEDENT/ \CONSEQUENT

Figure 7. A "partitioned" semantic network representing the sentence, "Every person owns a car." (If anode occurs in both the antecedent and the consequent space of an implication, it represents a universallyquantified variable. ISA denotes set membership. Frorn "Encoding Knowledge in Partitioned Networks"by G. G. Hendrix, in Associative Networks: Representation and Use of Knowledge by Computers editedby N. V. Findler, 1979, New York: Academic Press Inc. Copyright 1979 by Academic Press Inc. Reprintedby permission.)

consider nine of the more important oneshere.

1. The theorist must select an appropriatesubset of a language to represent within thenetwork. This choice will determine the ex-pressive power needed by the network. Acrucial decision is whether to admit quantifiedassertions, such as "No one voted for Someof the candidates," Early theories did notattempt to cope with quantifiers; subsequenttheories admitted them but were unable torepresent them properly, because as Woods(1975) pointed out, they were directly linkedto noun phrases as though they were adjec-tives. One problem with this practice is thatit fails to specify the scope of quantifiers, andit is therefore impossible to represent thecorrect meaning of sentences containing morethan one quantifier. Thus, the sentence "Noone voted for some of the candidates" meansthat there are some candidates for whom noone voted. Thfe quantifier "some" has a widerscope than the quantifier "no one" ratherthan vice versa. Later theories certainly admitquantification, provided that the quantifiersrange over individuals, that is, the networks

are equivalent in power to the first-orderquantificational calculus developed in formallogic.

2. The theorist has to decide how muchexpressive power to put into the network andhow much to put into the interpretativeprocesses that operate on it. This decision isagain crucial for quantified assertions. Somenetwork theories place the interpretative bur-den of quantifiers on the processes that op-erate on the network (e.g., ACT). However,partitioned semantic networks of the sortdevised as an exercise in artificial intelligence(Hendrix, 1979) can represent quantified as-sertions more directly. Figure 7 shows such arepresentation for the sentence, "Every personowns a car." Another method that could beimplemented is to use numerical indexes onquantified noun phrases to specify their rel-ative scopes (see Johnson-Laird, 1970).Among psychological theories, there is stillno consensus about whether all sentences,including quantified assertions, should be di-rectly representable in the network, or byprocesses that act on the network, or by someadmixture of the two components. Thus, the


theorist devising a new system here lacksconstraints to guide its development, becauseno principles have yet been formulated todetermine what goes into the network andwhat goes into the processes.

Because networks have been developedwithout a principled division of labor betweenrepresentation and process, we suggest thatthe following constraint might be adopted.The information in the network should rep-resent what people remember about themeanings of sentences (i.e., it should encodethe propositions that are expressed by dis-course). Hence, the processes that operate onthe network should play the same sort ofsemantic roles for all sentences, and thereshould be no difference from one sentence toanother in terms of the type of informationthat is represented in the network. It followsthat quantified and other complex sentenceswill call for the use of partitioned networksor some other structures of comparable power(cf. Brachman, 1979).

3. The theorist has to decide on the sortsof configuration that can occur within net-works representing sentences. Some systemsare restricted to a binary division of linksfrom each node (e.g., HAM), and othersallow an arbitrary number of links from eachnode (e.g., the LNR system). In principle,the semantic representation of a sentencecould be made within a strictly binary net-work in which each node has only two linksemanating from it. Certain predicates in nat-ural language express relations between morethan two arguments (e.g., "John sold Mary adog for $10." However, links can representfunctions that map two-place relations intorelations with a greater number of arguments,and then such sentences can be accommo-dated, because a label on one of the links inthe initial binary division can denote such afunction (see Figures 1, 2, and 6 for variationson this theme). The same issue arises in thedevelopment of data base management sys-tems and might be resolved in a similar way(cf. Bilger, 1979).

4. The theorist will have to decide whetherto use labels on nodes as well as on links(e.g., Quillian) or to use labels only on links(e.g., ACT). This question is largely a matterof computational convenience. To representthe fact that poodles are dogs, it makes no

semantic difference, given the appropriateprocessing system, whether one uses the or-thodox convention:

ISAPoodle —> Dog

or does without node labels:

Poodle Subject0 < 0' O

ISAPredicate .Dog

O »O >O.

5. The previous decision is related to thequestion of whether all links in the networkshould have the same semantic force. AsWoods (1975) pointed out, there is often nosingle interpretation for the links in the sys-tem. Such links as:

SizePoodle —> Small

ISAPoodle —> Dog

AgentBite > Poodle

are semantically heterogeneous. Size denotesa function that yields a value for a givenargument, ISA denotes a relation betweentwo sets, and agent denotes part of the relationbetween a verb and its arguments. Fromwithin the framework of network theory,however, there is no need to impose a uniforminterpretation on links. Moreover, if therewere such a need, a simple strategem isavailable to the theorist: Each link could betreated as denoting a function, and links thatnormally express relations could be treatedas denoting functions that return the truthvalues of relations, depending on whether therelation holds between the relevant entities.

6. A decision must be made about whetherto use standard grammatical labels such assubject and object on the links for sentencenetworks (e.g., ACT), Case-grammar labelssuch as agent, instrument, and recipient (e.g.,the LNR system), or some other idiosyncraticset of labels (e.g., HAM). This issue cannotbe resolved within the framework of networktheory. If a theorist's aim is to account solelyfor intensional relations, then it makes little


difference which system is adopted providedthat it is used consistently.

7. What should be the set of semanticlabels that are used in the network? Again,there is no single uniform and generallyaccepted set of labels other than ISA, HAS,and the type-to-token label. Intensional rela-tions place few constraints, if any, on theparticular set of labels that is selected. Nev-ertheless, a choice of labels may run intoproblems (see Chaffin & Herrmann, 1984),and unfortunately semantic network theorycontains as yet no principles for constrainingthe set of primitives.

8. It is necessary to decide whether infer-ences should be made by decomposition intosemantic primitives within the network (e.g.,LNR) or rules of inference ("meaning pos-tulates") that operate on the network repre-sentation (e.g., ACT). Once again, this issuecannot be decided within the framework ofnetwork theory, and some theorists claim thatthe choice is immaterial (J. R. Anderson,1976). We argue later that networks fail torepresent the crucial information requiredfor inferences; it follows that neither approachis likely to be completely successful.

9. Performance in evaluating intensionalproperties and relations should be explicablein terms of computations carried out on thenetwork (e.g., the latency of a response shouldreflect such factors as the distance that hasto be traversed within the network, the orderin which links are searched, and the natureof the subject's task). The theorist has todecide how these processes are to operate.However, there appear to be no constraintson the sort of processes that can be invokedto model performance in semantic tasks.Typically, processing calls for a search for apath from one node to another, but compu-tations on the information represented in thenetwork could take any form that the theoristdesires. Indeed, most network theories seemto have the power of a Universal Turingmachine (i.e., the processes they invoke aresufficient to compute anything that can becomputed at all), granted Turing's thesis—yet to be falsified—that any function forwhich there is an effective procedure can becomputed by a Turing machine (see, e.g.,Rogers, 1967). J. R. Anderson (1976) has infact shown that the ACT system is equivalent

in power to a Universal Turing machine, anddoubtless the equivalence could be provedfor other network systems, especially thosethat employ augmented transition network(ATN) parsers (e.g., LNR),

These nine decisions face anyone develop-ing a network theory. They constrain thenature of the resulting system to some extent,and we have suggested some additional factorsthat should be considered in making thedecisions. Yet the constraints remain too fewand too weak; current network systems aretoo powerful. Any particular network theoryusually yields clear empirical predictions, butthe class of network theories is not con-strained: If networks can compute anything,then plainly as a class of theories they arealmost empirically vacuous. The moral ofTuring machine equivalence is that any ex-perimental result that cannot be accountedfor by an existing network can always beaccommodated by appropriate revisions.Similarly, any other theory of intensionalphenomena can be reexpressed as a networktheory. Thus, theories based on semanticfeatures (see, e.g., McCloskey & Glucksberg,1979; Schaeffer & Wallace, 1970; Smith etal., 1974) or meaning postulates (e.g., Fodoret al., 1975; Kintsch, 1974) can be readilytranslated into networks (see also Hollan,1975). A major problem is precisely the lackof a general theory that restricts the class ofpotential networks to a readily testable subset.We have no quarrel with the formalism ornotation of networks: A commitment to themis little more restrictive, and no more opento criticism, than is a commitment to aparticular programming language such asLISP.

We have now described the main varietiesof semantic networks, outlined their commonunderlying assumptions and their points ofdifference, and showed that the class of the-ories as a whole is insufficiently constrained.Our next task is to assess network theory inrelation to the four major goals of semantictheory that we presented in the first sectionof the article.

Some Problematical Semantic Phenomena

The most striking design feature of seman-tic networks is that they have virtually nothing


to say about extensional relations. In contrast,model-theoretic semantics, which derives fromwork in formal logic, provides a method ofspecifying the extensions of expressions withina model structure—typically, a set of numbersor some other abstract domain. The theoristposits the extensions of the basic lexical itemsin the model structure and formulates a setof semantic rules (which usually work intandem with the syntactic rules of the lan-guage) for combining these extensions toform the extensions of expressions, and soon, all the way up to the truth value of thesentence. The method has been applied tonatural language by Montague (1974) andothers.

Some network theorists have recognizedthat networks lack any extensional machinery,and J. R. Anderson (1976) has even putforward a model-theoretic semantics for hisACT system that specifies the truth conditionsof sentences in terms of the extensions oftheir constituents, Anderson's aim, unfortu-nately, is not to formulate a theory of howpeople represent extensional relations, butrather to use this semantics to determine theexpressive power of the ACT system.

Most of the problems with current semanticnetworks arise from the neglect of extensionsand from the assumption that the intensionalrelations between expressions can be analyzedindependently from extensional matters. Inthis section, we are going to consider fivephenomena: the relations among intensionalrelations, ambiguity, anomaly, instantiation,and inference. Any psychological theory ofmeaning should account for these phenomena;semantic networks contain mechanisms de-.signed to do so, but nevertheless fail to dealwith them adequately, a failing that alsoapplies to theories based on semantic featuresor on meaning postulates.

Relations Among Intensional Relations

Although semantic networks are designedto represent intensional relations, they arenot readily able to accommodate the fullrange of phenomena associated with them.As network theory acknowledges, there aredifferent types of intensional relation (e.g.,synonymy, antonymy, class inclusion, and

partonymy). Moreover, native speakers canmake judgments about types of intensionalrelation. They can judge, for example, thatlarge is opposite to small, that slim is similarto thin, that silver contrasts with gold, andthat a kitchen is a part of a house. Networktheory can explain these judgments by usingadditional labelled links. For example, thewords large and small might be joined by alink labelled as antonymous (see Glass, Hol-yoak, & Kiger, 1979). Alternatively, theymight each be linked to a higher order node,such as size, with these links marked asinconsistent (Glass & Holyoak, 1974/1975).However, such links account for neither thepatterns of similarity among the differenttypes of intensional relation nor subjects'judgments about these patterns. Thus, syn-onymy (e.g., car-auto) is judged as moresimilar to class inclusion (car-vehicle) thanit is to partonymy (car-engine). These differ-ences in similarity are reliable across tasksand studies spanning several decades (Chaffin& Herrmann, 1981, 1984; Chaffin, Winston,& Herrmann, 1984; Moran, 1941; Perfetti,1967; Riegel & Riegel, 1963; Warren, 1921).

Types of intensional relation can, of course,be analyzed and defined (see, e.g., Evens,Litowitz, Markowitz, Smith, & Werner, 1983),and subjects' judgments about the patternsof similarity can be explained in terms oftheir having access to some such underlyingspecifications (Chaffin, Russo, & Herrmann,1981; Herrmann et al., 1979). Likewise, thelatencies of their judgments about the inten-sional relation between a pair of words canbe accounted for by the degree to which theactual relation between the pair conforms tothe specification of the type of intensionalrelation. For example, judgments that twowords are antonymous are faster if they aresymmetrically opposed (e.g., large-small) thanif they are asymmetrically opposed (e.g.,large-tiny; Herrmann, Chaffin, & Daniel,1984). Judgments that two words are synon-ymous are faster if they have the same setsof extensions (e.g., car-auto) than if the setsare merely similar (e.g., taxi-limousine;Herrmann, 1978; Herrmann, Papperman, &Armstrong, 1978). There are similar resultswith set inclusion (Loftus, 1973; Meyer, 1970)and part-whole judgments (Chaffin et al.,1984). Because current networks do not rep-


resent the underlying specifications of inten-sional relations, which are treated as unana-lyzed labels on links, they cannot accountfor judgments about types of intensional re-lation.

There are various ways in which the re-quired machinery might be introduced intonetworks. The most natural method wouldbe to represent the underlying semantics ofthe different types of intensional relationwithin the network itself. It would be neces-sary to formulate definitions of synonymy,antonymy, and the other types of relation interms of more primitive notions and then toconstruct network representations of thesedefinitions. For instance, one component ofantonymy is that the two terms should denoteopposite values of the same property; tall andsmall are not antonyms because the formerapplies to height and the latter, to size. Like-wise, if two terms deal with an underlyingcontinuum, then they should be symmetricabout the midpoint; hot and cold are anto-nyms, whereas hot and tepid are not. Similarintensional relations would obviously havesimilar representations in the network. Thispattern would account for subjects' judgmentsof similarity. The degree to which the linksbetween a pair of words conformed to thedefinition of a given intensional relation wouldlikewise account for the latencies of suchjudgments (Herrmann & Chaffin, 1984).What is unclear, however, is the nature of thelinks required to define the various types ofintensional relation and the nature of theprocesses for comparing definitions with theactual links between pairs of words. Giventhe power of networks, we have no doubtthat such machinery could be developed, buteven this addition leaves further phenomenaunexplained.

In particular, there are analogous problemsthat arise with relations that hold, not betweenthe senses of words, but between what theyare actually used to refer to—their extensions.For example, anyone who has ever packedgoods at a supermarket is familiar with thefact that some things are more easily squashedthan others. Tomatoes are more squashablethan potatoes. Semantic networks are sup-posed to be appropriate representations forall concepts, but no feasible network couldcontain a label between tomatoes and potatoes

that expresses this relation (Chaffin & Herr-mann, 1984). Moreover, if one were to ignorethe exponential growth in the number oflinks that would result from adding suchlinks, then there is still a problem. If potatoesare cooked and mashed, or if tomatoes arefrozen solid, then it is no longer the case thattomatoes are more squashable than potatoes.Judgments about such relations do indeedrequire access to the extensions of expressions.

Other relations cause problems becausethey are ad hoc and more abstract or complex.They include Boolean functions of simplerelations, like the set of all pairs, X and Y,such that X is larger and faster then Y, andquantified relations like the set of all pairs ofsets, X and Y, such that some members of Xare taller than all members of Y. The latterhas quantifiers that range over the individualsin two sets (e.g., the relation holds betweenmen and women, adults and children, animalsand human beings, but it does not holdbetween the converse pairs). Because only afraction of all possible Boolean and quantifiedrelations would be directly represented in anetwork, the evaluation of a relation wouldrequire in most cases a computation thatcompared the range of values associated withthe two terms. For example, the verificationof "An airplane is larger and faster than asquirrel" could not be done by a searchthrough the network because there would beno direct links labelled "larger than" and"faster than" connecting the two terms. Itwould therefore be necessary to compare therespective sizes and speeds associated withairplane and squirrel and to determinewhether the differences between them satisfiedthe sense of "larger and faster than."

Of course, the existence of a mechanismfor evaluating the definitions of intensionalrelations does not rule out the possibility thatsome intensional relations are directly rep-resented in a network (or some other formof intensional representation, such as a set ofsemantic features or meaning postulates).What relations might be encoded in this way?A plausible answer is those relations that aredirectly learned in a metalinguistic fashion,When a child learns that little is the oppositeof big, or that large is similar in meaning tobig, then these metalinguistic facts could bedirectly represented by labelled links between


the representations of the appropriate words.Most intensional relations, however, are notdirectly acquired for all the terms that satisfythem (e.g., children do not learn the set ofopposites exhaustively). Likewise, many in-tensional relations themselves are never ex-plicitly learned. It follows that the directrepresentation of intensional relations is likelyto play only a small part in semantic judg-ments.

Ambiguity

Comprehensive network theories are de-signed to represent the meanings of sentences.They therefore have to be able to cope withthe resolution of ambiguities. Many wordsare in fact ambiguous, and the more fre-quently used words are more likely to beambiguous (see Miller, 1951). Speakers andlisteners seldom notice these ambiguities un-less the sentence as a whole is ambiguous,and hence the mechanisms for resolving lex-ical ambiguity must be highly efficient. Se-mantic networks cope with ambiguous wordsby using so-called "selectional restrictions"to determine their appropriate interpretations.This idea derives from a standard procedurein lexicography: Dictionaries relate each dif-ferent sense of a word to the meanings ofother words that can occur in constructionwith it. The meanings of these other wordsaccordingly select the appropriate sense asshown in the following example:

plant (v.t.): to set or sow (seeds or plants)in the earth;

to establish (animals) in a newlocality,

where the parenthetical items in the definitionare the selectional restrictions. Katz and Fodor(1963) developed this approach into a moreformal theory, and Quillian (1968) and hisintellectual descendants adopted it as well.An ambiguous word is disambiguated byestablishing the words that occur in construc-tion with it in the same sentence and checkingwhich restrictions stated in its network defi-nition their meanings satisfy. Hence, the sen-tence, "He planted his herd on the island,"cannot mean that he sowed the earth withanimals because this sense of the verb isrestricted to seeds or plants.

Unfortunately, selectional restrictions haveonly a limited explanatory value. What isneeded in disambiguating a word is access tothe extensions of other words in the sentence.This point becomes clear from consideringhow the verb in the following sentence isinterpreted: "He planted them on the island."If, for instance, the pronoun "them" refersto seeds, then the verb will be taken to meanthat the seeds were sown. However, the pro-noun does not have seeds as one of itsmeanings; rather, it is the extension of theterm. At the very least, therefore, the appa-ratus of selectional restrictions embodied insemantic networks must be modified so thatit has access to extensions.

There is a still more serious problem. It isimpossible to state complete and consistentselectional restrictions on the meanings ofmany ambiguous words. Consider, for in-stance, the transitive verb "to lift," which hasa number of different meanings, including:to move (an object) from a lower to a higherposition; to steal or plagiarize (an idea); andto put an end to (a siege or blockade orembargo). What restrictions does the firstsense place on the subject of a sentence ofthe form "X lifted F"? As a first approxi-mation, one might suppose that X should behuman, animal, or a machine. However, thishypothesis fails to accommodate the followingsentence: "The wind lifted the leaves." Itseems the selectional restriction should be:capable of exerting a physical force, and thatevery item in the network denoting a human,animal, machine, or entity such as the windshould be linked to a node representing thismeaning. However, such a scheme will stillnot work properly. Further machinery isneeded to cope with the fact that the particularentity referred to by the subject of the sen-tence may, or may not, be capable of exertinga force depending on the circumstances towhich the assertion refers. Thus, for example,although human beings are generally capableof exerting a force, dead or paralyzed humanbeings are not.

A sensible construal of the theory of selec-tional restrictions is that it captures thoseinferences that are made so often that theyhave become assumptions about how wordsare to be interpreted by default (see Miller &Johnson-Laird, 1976, p. 701). Hence, unless


there is evidence to the contrary, the subjectof "lift," if it is to be interpreted as meaningmove upward, should be animate or machine-like. In general, however, lexical ambiguitiescan be resolved only by inferences based onknowledge about the specific entities referredto in the sentence.

Anomaly

Semantic anomaly arises whenever a sen-tence cannot be given a meaningful interpre-tation. For example, the sentence, "The bowlof soup ate the ham sandwich" on any feasibletheory of selectional restrictions ought to beanomalous, because only animate entities caneat sandwiches. The sentence might, of course,have the metaphorical interpretation that theham sandwich fell into the bowl of soup andwas engulfed by it; and this interpretationmight be explained by mechanisms that comeinto play only when the system detects asemantic anomaly. As Bolinger (1965) wrote,"It is characteristic of natural language thatno word is ever limited to its enumerablesenses, but carries with it the qualification,'something like'." However, there is quite an-other interpretation of the sentence abovethat is not metaphorical. Nunberg (1978)pointed out that such noun phrases as "thebowl of soup" and "the ham sandwich"could be readily used by a waiter as a way ofreferring to his customers in terms of whatthey have ordered. A theory that assumesthat a set of literal meanings is first assignedto a sentence and that context then eliminatesinappropriate meanings runs into grave dif-ficulties with such sentences. Once again, theproper treatment of this intensional phenom-enon calls for semantic networks to haveaccess to extensions.

Instantiation

The linguistic context in which an unam-biguous word occurs can narrow down itsinterpretation. This phenomenon of "instan-tiation" has been demonstrated experimen-tally by R. C. Anderson and his colleagues.They presented subjects with a sentence suchas "The fish attacked the swimmer," and laterthe subjects had to recall the sentence (seeR. C. Anderson et al., 1976). A more specificand likely term such as shark was a better

recall cue than the original general term, fish.Analogous phenomena occur with the inter-pretation of verbs (Garnham, 1979), andcontext can also render one aspect of themeaning of a word more salient than another(Tabossi & Johnson-Laird, 1980).

Halff, Ortony, and Anderson (1976) ac-counted for instantiation by arguing thatwords do not have a few qualitatively distinctmeanings, but rather a whole family of po-tential meanings (see also Putnam, 1975;Weinreich, 1966). When a word occurs in asentence, the linguistic context acts to instan-tiate a specific member of the family ofmeanings. This argument is mildly embar-rassing for network theories, because it impliesthat each word should be associated withmany more concept nodes than theoristsnormally envisaged. However, there is analternative explanation of instantiation thatis both more convincing and more embar-rassing for network theories. The sentence,"It attacked the swimmer," might well bebetter recalled given the cue shark ratherthan the original term, it. No one wouldseriously argue that "it" is polysemous; it hasa single meaning, which enables it to refer toan infinite number of different entities. Hence,what is instantiated by linguistic context isnot a particular sense but a particular referent.There is no need to invoke vast sets ofmeanings for words, but there is a need foraccess to extensions.

Inference

One function of semantic networks is toenable inferences to be drawn from verbalassertions. For example, the deduction, "Fidois a poodle; therefore, Fido is a dog" dependson the unstated premise that all poodles aredogs. The inference can be made by traversingfirst the link that establishes that Fido is amember of the set of poodles and then thelink that establishes that the set of poodles isincluded in the set of dogs. This method ofmaking inferences is formal in that oneexpression is derived from others by virtueof a configuration of symbols and withoutregard to their extensions or truth conditions.Indeed, J. R. Anderson argued for the networkrepresentation of HAM in part because itsuggested this sort of inferential mechanism


(cf. J. R. Anderson, 1976, p. 41). However,there are valid inferences that cannot bemade in a formal way, because the interpre-tation of terms depends on the situationreferred to rather than on the senses of theexpressions. For example, inferences of theform, A is on B's right; B is on C's right;therefore, A is on C's right, are valid providedthat A, B, and C are seated down one side ofrectangular table and facing in the samedirection. The inference may be invalid if therelevant individuals are seated around a cir-cular table. It is of little use to claim that theexpression "on X's right" has two meanings—one transitive and one intransitive—becausewith a seating arrangement of an appropriateradius transitivity can extend over some ar-bitrarily finite number of individuals (John-son-Laird, 1981). These vagaries of deductiveinference require an inferential system thathas access to a representation of the situationto which reference is made.

Extensional Models and Redundancy ofIntensional Relations

Symbolic Fallacy

In the previous section, we outlined someproblems that arise if intensional relationsare supposed to be interpreted independentlyfrom extensions. Semantic networks are, ofcourse, based on that assumption. Indeed,theorists have assumed that the process oftranslating natural language into a networkrepresentation amounts to comprehension.They have overlooked the fact that such arepresentation is strictly meaningless unlessin principle it can be connected to a concep-tion of the external world; and, of course,network theory does not establish any suchconnection. In this respect, semantic networksrender theorists particularly vulnerable towhat we call the "symbolic" fallacy—theassumption that the mere translation of sen-tences into symbols constitutes a useful ac-count of their meaning (see also Lewis, 1972).For example, Kintsch and van Dijk (1978)proposed a propositional representation forsentential meanings, and they were sensitiveto the potential emptiness of the procedure.When they translated "The students com-plain" into (complain, student), they pointedout:

Complain is merely the name of a knowledge complexthat specifies the normal uses of this concept, includingantecedents and consequences. Inferences can be derivedfrom this knowledge complex, for example, that thestudents were probably unhappy or angry or that theycomplained to someone, with the text supplying theprofessor for that role. Once a concept like complain iselaborated in that way, the semantic notation is anythingbut vacuous. (Kintsch & Van Dijk, 1978, p. 378)

Unfortunately, if the knowledge complexmerely consists of other symbolic expressions,then there is no escape from the maze ofsymbols into the world: The theory has suc-cumbed to the symbolic fallacy. The fallacyis analogous to the hackneyed science fictionidea that the aliens learned the languages ofthe Earth's inhabitants from radio transmis-sions. One might (just) learn that "zugbrochna" follows from "gek brochna" in thisway, but no matter how complex the infer-ences that can be acquired, the system willnever amount to a proper semantics. It con-nects expressions to expressions perhaps, butit does not connect them to their extensions.This issue is precisely the one that croppedup in different guises with the problematicalphenomena of the previous section: Theycould not be handled by networks, whichlack representations of the extensions ofexpressions. Of course, network theories arenot alone in committing the symbolic fallacy:A reluctance to deal with reference is evidentin theories based on semantic features or onmeaning postulates.

Models of Extensions

Is it possible to remedy network theory sothat it deals with extensions and thus copeswith the puzzling aspects of intension rela-tions? The answer is both yes and no. Anyputative theory can be reexpressed in theform of a network system because, as wementioned earlier, networks in general havethe power of Universal Turing machines. Incontrast, as we show, the required solutionappears to call for structures that are foreignto network theory, and so the revised networkswill not look at all like the sorts of theorythat we have considered in this review.

A fruitful way in which to approach theproblem of extensions is to develop a psycho-logically oriented model-theoretic semanticsbased on the assumption that sentences relate


to the world by virtue of mental models ofstates of affairs (Johnson-Laird, 1983). Hu-man beings can construct models of theworld on the basis of perception, memory,and imagination; they can also constructrudimentary models of the world on the basisof verbal descriptions; and they can sometimesevaluate an expression by comparing its lin-guistically derived model with a perceptualor conceptual one. In many cases, of course,the process of comparison may be difficult:If the truth conditions concern social relationsor states of affairs that have no direct percep-tual correlate in the world, then the relevantevidence may be at best indirect. In othercases, the process of comparison may beimpossible: If the truth conditions of theassertion are incomplete or indeterminate,then no evidence may suffice to verify theassertion. Nevertheless, the fact that humanbeings can in principle relate language tomodels of the world provides the foundationof semantics.

A semantic theory for mental models re-quires two levels of representation. First, anutterance has to be translated into an inten-sional representation—a representation of theproposition that the utterance expresses. Sec-ond, the prepositional representation may betranslated into an extensional representa-tion—a mental model of the particular stateof affairs characterized by the utterance. Thetwo levels are required in order to accountfor a number of inferential and interpretativephenomena, and they are borne out by ex-perimental evidence (e.g., Mani & Johnson-Laird, 1982; Ehrlich & Johnson-Laird, 1982).Most important, they allow for the recursiverevision of the model if, as in the case of anindeterminate description, it should embodysome assumption that subsequently turns outto be erroneous. In this way, a single specificmodel can represent a proposition: The sys-tem sets up a representative sample from theset of possible models of the discourse, andif the sample fails to match a subsequentassertion, it can be revised (within the con-straints of human memory) so as to beconsistent with the discourse as a whole. Thisapproach solves the problem of how a singlemodel of, for example, a cat sitting on a matcan represent the indefinite number of differ-ent states of affairs that are truthfully de-

scribed by a sentence to that effect. Theprocedures that revise the model, however,must have access to an independent recordof the sentence, and the initial prepositionalrepresentation serves this function.

The prepositional representation of a sen-tence may well take the form of a semanticnetwork, but plainly there is a need for afurther extensional representation, or mentalmodel, of the state of affairs described by thesentence. Mental models, unlike semanticnetworks, have a structure that correspondsto the perceived or conceived structure ofthat state of affairs. Thus, for example, amental model of an assertion such as "Everyperson owns a car" contains an arbitrarynumber of tokens corresponding to persons,an arbitrary number of tokens correspondingto cars, and relations between these two setsof tokens designating ownership:

Person 1 —» Car 1

Person 2 —» Car 2

Person 3 -> Car 3

(Car 4)

The parenthetical item represents the factthat there may be cars that are not owned byanyone. It is important to note that thestructural constraint on mental models (i.e.,that their structure corresponds to the per-ceived or conceived structure of states ofaffairs) renders them quite distinct from con-ventional semantic networks. Thus, the net-work representation of the same sentence,which is shown in Figure 7, calls for fourpartitions, eight nodes, and eight links—astructure that is remote from the structureof an actual state of affairs in which everyperson owns a car, where indeed there willbe two sets with a relation from all themembers of one set to members of the other.

If mental models are to be constructedfrom sentences, the meanings of words haveto specify the conditions that must be satisfiedin the model for the word to apply truthfullyto it. The lexical semantics must thereforecapture the contribution that the word makesto the truth conditions of any assertion inwhich it occurs. Thus, for example, the se-mantics for the expression "on the right of"specifies the direction to be scanned in order


to form a mental model of such assertions as"A is on the right of B." The interpretativesystem locates one entity within a spatialarray and then scans in an appropriate direc-tion with respect to the observer's viewpointin order to locate the other. The directioncan be specified straightforwardly (e.g., usingiteration on Cartesian coordinates), but ob-viously such a specification calls for conceptsthat may well be ineffable in the languageunder analysis. Indeed, one cannot define thetruth conditions for the sentence "A is on theright of 5" in ordinary language.

The conjecture that subjects constructmental models of the events described indiscourse goes a considerable way towardexplaining the resolution of ambiguities,anomalies, instantiation, the vagaries of in-ference, and the well-known findings ofBransford and his colleagues that subjectsoften go beyond what is linguistically givenin interpreting discourse (Bransford & Mc-Carrell, 1975).

Specifications of Intensional RelationsMay Be Redundant

If a computer program is based on thetheory of mental models (see Johnson-Laird,1983), a striking consequence emerges: It canwork without relying on any representationof intensional relations of the sort capturedin semantic networks. Thus, the program canmake transitive inferences without recourseto rules of transitive inference. For example,the truth conditions for the assertion "A ison the right of B" enable an appropriatemodel to be constructed: B A. Likewise, thetruth conditions for the further assertion "Bis on the right of C" enable the program toextend the model appropriately to: C B A.The program contains a verification routinethat is elicited whenever an assertion makesno reference to any new entities. The sametruth conditions for "on the right of" there-fore suffice to return the value, "true", whena third assertion is made: "A is on the rightof C." At this point, of course, no validdeduction has been made. However, the pro-gram also contains a recursive procedure thatsearches for an alternative model of thepremises that would render the current asser-tion false. There is no such alternative in this

case, and (the program concludes that) thededuction is valid. In summary, the programmakes a transitive inference without needingany rule to the effect that "on the right of"is transitive: Transitivity is an emergent prop-erty from the truth conditions of the relation.

The simplest way in which to state theredundancy of intensional relations is in termsof the following example. If you know whatit is for something to be a canary and youknow what it is for something to be a bird,then you do not need to know explicitly thatcanaries are birds, because this relation willbe a consequence of your knowledge. Ofcourse, the fact that information is redundantdoes not imply that it is not mentally repre-sented.

We have reached the same conclusion thatwe arrived at in considering relations amongintensional relations. Some intensional rela-tions will be mentally represented, others willnot. Those that are not mentally representedcan always be recovered on the basis ofprocesses that examine the truth conditionsof terms. Which relations are likely to beredundantly specified? The answer, onceagain, is those that are directly learned. Hence,if you learn that canaries are birds, then thisrelation will be mentally encoded even thoughyou may also learn about the extensions ofthe two terms. In contrast, there are termsfor which you initially acquire a direct, al-though ineffable, knowledge of truth condi-tions. These terms are likely to include whatRosch (1976) called "basic-level" words, thatis, words that denote entities that have attri-butes and functions maximally in common.

Conclusions

How do semantic networks fare in relationto the four goals for a psychological theoryof meaning that we outlined earlier? Theyspecify a form for the mental representationof meaning (the first goal), although this formis remote both from the immediate structureof the sentence and from the conceived struc-ture of the world. They also account for someintensional properties and relations (the sec-ond goal), but they run into difficulties withcertain cases of relations between words. Theyare not designed to specify the extensions ofwords or expressions (the third goal). In


addition, their various accounts of inference(the fourth goal), whether by decompositionor by rules, are not entirely adequate forcoping with the vagaries of validity.

We can state our misgivings about semanticnetworks by way of an example. One of thethings that you know about poodles is thatthey are dogs, and this sort of informationcan be represented in a semantic network.However, you also have some knowledge aboutwhat it is for something to be a poodle; youhave a concept of what poodles are, and thisknowledge enables you to identify poodles,to establish complex intensional relations in-volving them, and to verify assertions aboutthem. Of course, no one can definitivelyclassify any entity as either a poodle or nota poodle: There are probably no necessaryand sufficient conditions for poodlehood (seePutnam, 1975), but without some knowledgeof what determines the extension of a termyou can hardly be said to have grasped itsmeaning. This sort of knowledge is not rep-resented in semantic networks, and there isno immediate way in which it could berepresented because networks lack connec-tions to representations of the world. Theyonly provide connections between words. Un-fortunately, they cannot even give a completeexplanation of intensional phenomena, be-cause a proper account of ambiguity, anomaly,instantiation, and inference turns out to de-pend on access to extensional representations.Because, as we noted earlier, network theoryis a notational variant both of models basedon semantic features (e.g., McCloskey &Glucksberg, 1979; Schaeffer& Wallace, 1970;Smith et al., 1974) and of models based onmeaning postulates (e.g., Fodor et al., 1975;Kintsch, 1974), the shortcomings applyequally to these theories.

When a theory is constructed to handleextensions, then it becomes clear that it is nolonger strictly necessary to represent the in-tensional connections between words. Anysuch theory that accommodates extensionscould undoubtedly be reexpressed within theformalism of a network system. But thistranslation would be very different from cur-rent network theory. The new theory wouldbe serving an extensional function, not ad-dressed by previous network theories, andthe structures of its network would be radi-

cally different because they would now cor-respond to the structures with which humanbeings conceive states of affairs.

Despite these strong qualifications on net-work'theory, it remains highly likely thatcertain intensional relations are directlylearned as such, and that they may accordinglybe represented by associative relations be-tween representations of words. Semanticnetworks may therefore still have a role,albeit a reduced one, in accounting for themental representation of the meanings ofwords. They may also have a major role as ahypothesis about the structure of the initialencoding of sentences especially if they aresupplemented by some form of extensionalrepresentation. The moral of this article issimple. Semantic networks can do manythings, but they cannot do everything that apsychological theory of meaning ought to beable to do: The meanings of words can onlybe properly connected to each other if theyare properly connected to the world.

References

Anderson, J. R. (1976), Language, memory, and thought.Hillsdale, NJ: Erlbaum.

Anderson, J. R., & Bower, G. H. (1973). Human associa-tive memory. Washington, DC: Winston.

Anderson, R. C., Pichert, J. W., Goetz, E. T., Schallert,D. L., Stevens, K. V,, & Trollip, S. R. (1976). Instan-tiation of general terms. Journal of Verbal Learningand Verbal Behavior, 15, 667-679.

Bierwisch, M. (1970). Semantics. In J. Lyons (Ed.), Newhorizons in linguistics (pp. 166-184). Harmondsworth,Middlesex, England: Penguin.

Bilger, H. (1979). On the notion of irreducible relations.In G. Bracchi & G. M, Nijssen (Eds,), Data basearchitecture (pp. 277-296). Amsterdam: North-Holland.

Bobrow, D. G., & Collins, A. M. (Eds.). (1975). Repre-sentation and understanding: Studies in cognitive sci-ence. New York: Academic Press.

Bolinger, D. (1965). The atomization of meaning. Lan-guage. 41, 555-573.

Brachman, R. J. (1979). On the epistemological statusof semantic networks. In N. V. Findler (Ed.), Associativenetworks: The representation and use of knowledge bycomputers (pp. 3-50). New York: Academic Press.

Bransford, J. D., & McCarrell, N, S. (1975). A sketch ofa cognitive approach to comprehension: Some thoughtsabout understanding what it means to comprehend. InW. B. Weimar & D. S. Palermo (Eds.), Cognition andthe symbolic processes (pp. 189-230). Hillsdale, NJ:Erlbaum.

Chaffin, R., & Herrmann, D. J. (1981). Semantic rela-tionships and the generality of semantic-memory mod-els. Bulletin of the Psychonomic Society, 12, 69-72.

Chaffin, R., & Herrmann, D. J. (1984). The diversity


and similarity of semantic relations. Memory andCognition, 12, 134-141.

Chaffin, R., Russo, A., & Herrmann, D. J. (1981). Aneffect of relationship similarity on categorization latency.Journal of General Psychology, 104, 305-306.

Chaffin, R., Winston, M., & Herrmann, D. J. (1984).The semantics of part-whole relations: An empiricaltaxonomy. Unpublished manuscript.

Chomsky, N. (1965). Aspects of the theory of syntax.Cambridge, MA: MIT Press.

Clark, H. H., & Clark, E. V. (1977). Psychology andlanguage: An introduction to psycholinguistics. NewYork: Harcourt Brace Jovanovich.

Collins, A. M., & Loftus, E. F. (1975). A spreading-activation theory of semantic processing. PsychologicalReview, 82, 407-428.

Collins, A. M., & Quillian, M. R. (1969). Retrieval timefrom semantic memory. Journal of Verbal Learningand Verbal Behavior, 8, 240-247.

Collins, A. M., & Quillian, M. R. (1972). How to makea language user. In E. Tulving & W. Donaldson (Eds.),Organization and memory (pp. 310-349). New \brk:Academic Press.

Conrad, C. (1972). Cognitive economy in semantic mem-ory. Journal of Experimental Psychology, 92, 149-154.

Deese, J. (1962). On the structure of associative meaning.Psychological Review, 69, 161-175.

Deese, J. (1965). The structure of associations in languageand thought. Baltimore, MD: The Johns HopkinsUniversity Press.

Ehrlich, K., & Johnson-Laird, P. N. (1982). Spatialdescriptions and referential continuity. Journal of VerbalLearning and Verbal Behavior, 21, 296-306.

Evens, M. W., Litowitz, D. E., Markowitz, J. A., Smith,R. N., & Werner, D. (1983). Lexical-semantic relations:A comparative survey. Current inquiry into languageand linguistics (Report No, 34). Carbondale, IL andEdmundton, Alberta, Canada: Linguistic Research Inc.

Fillmore, C. J. (1968). The case for case. In E. Bach &R. T. Harms (Eds.), Universals in linguistic theory (pp.1-88). New York: Holt, Rinehart & Winston.

Findler, N. V. (Ed.). (1979). Associative networks: Rep-resentation and use of knowledge by computers. NewYork: Academic Press.

Fodor, J. D., Fodor, J. A., & Garrett, M. F. (1975). Thepsychological unreality of semantic representations.Linguistic Inquiry, 4, 515-531.

Frege, G. (1892). (Jber sinn und bedeutung [On senseand reference]. Zeitschrift fur Philosophie und philo-sophische Kritik. 100, 25-50. In P. T. Geach & M.Black (Eds. and Trans.), Translations from the philo-sophical writings of Gottlob Frege. Oxford: Blackwell,1952.

Garnham, A. (1979). Instantiation of verbs. QuarterlyJournal of Experimental Psychology, 31, 207-214.

Glass, A. L., & Holyoak, K. J. (1974/1975). Alternativeconceptions of semantic memory. Cognition, 3, 313-339.

Glass, A. L., Holyoak, K. J., & Kiger, J. I. (1979). Roleof antonymy relations in semantic judgments. Journalof Experimental Psychology: Human Learning andMemory, 5, 598-606.

Halff, H. M., Ortony, A., & Anderson, R. C. (1976). Acontext-sensitive representation of word meanings.Memory and Cognition, 4, 378-383.

Hendrix, G. G. (1979). Encoding knowledge in partitionednetworks. In N. V. Findler (Ed.), Associative networks:Representation and use of knowledge by computers(pp. 51-92). New York: Academic Press.

Herrmann, D. J. (1978). An old problem for the newpsychosemantics: Synonymity. Psychological Bulletin,55,490-512.

Herrmann, D. J., & Chaffin, R. (1984). Definitions ofsense relations. Unpublished manuscript.

Herrmann, D. J., Chaffin, R., & Daniel, M. (1984).Three attributes of antonymy: Effects of relation attri-butes on antonym comprehension. Manuscript submit-ted for publication.

Herrmann, D. J., Chaffin, R. J. S., Conti, G., Peters, D.,& Robbins, P. H. (1979). Comprehension of antonymyand the generality of categorization models. Journal ofExperimental Psychology: Human Learning and Mem-ory, 5, 585-597.

Herrmann, D. J., Papperman, T. J., & Armstrong, A. D.(1978). Synonym comprehension and the generality ofcategorization models. Memory and Cognition, 6, 150-155.

Hollan, J. D. (1975). Features and semantic memory:Set-theoretic or attribute model? Psychological Review,82, 154-155.

Holyoak, K. J., & Glass, A. L. (1975). The role ofcontradictions and counterexamples in the rejectionof false sentences. Journal of Verbal Learning andVerbal Behavior, 14, 215-239.

Humphrey, G. (1951). Thinking: An introduction to itsexperimental psychology. London: Methuen.

Johnson-Laird, P. N. (1970). The interpretation of quan-tified sentences. In G. B. Flores d'Arcais & W. J. M.Levelt (Eds.), Advances in psycholinguistics (pp. 347-372). Amsterdam: North-Holland.

Johnson-Laird, P. N. (1974). Experimental psycholin-guistics. Annual Review of Psychology, 25, 135-160.

Johnson-Laird, P. N. (1981). Mental models of meaning.In A. K. Joshi, B. L. Webber, & I. A. Sag (Eds.),Elements of discourse understanding (pp. 106-126).Cambridge, England: Cambridge University Press.

Johnson-Laird, P. N. (1983). Mental models: Towards acognitive science of language, inference, and conscious-ness. Cambridge, England: Cambridge University Press.

Katz, J. J. (1972). Semantic theory. New York: Harper &Row.

Katz, J. }., & Fodor, J. A. (1963). The structure of asemantic theory. Language, 39, 170-210.

Kintsch, W. (1974). The representation of meaning inmemory. Hillsdale, NJ: Erlbaum.

Kintsch, W, & van Dijk, T. A. (1978). Towards a modelof text comprehension and reproduction. PsychologicalReview, 85, 363-394.

Kiss, G. R. (1967). Networks as models of word storage.In N. L. Collins & D. Michie (Eds.), Machine intelli-gence, Vol. /(pp. 155-168). Edinburgh: Oliver & Boyd.

Landauer, T. K., & Freedman, J. L. (1968). Informationretrieval from long-term memory: Category size andrecognition time. Journal of Verbal Learning andVerbal Behavior, 7, 291-295.

Lewis, D. K. (1972). General semantics. In D. Davidson& G. Harman (Eds.), Semantics of natural language(pp. 169-218). Dordrecht, Holland: Reidel.

Lindsay, P. H., & Norman, D. A. (1977). Human infor-


motion processing: An introduction to psychology (2nded.). New York: Academic Press.

Loftus, E. F. (1973). Category dominance, instance dom-inance, and categorization time. Journal of Experi-mental Psychology, 97, 70-74.

Lyons, J. (1977). Semantics (Vol. 1). Cambridge, England:Cambridge University Press.

Mani, K., & Johnson-Laird, P. N. (1982). The mentalrepresentation of spatial descriptions. Memory andCognition, 10, 181-187.

McCloskey, M., & Glucksberg, S. (1979). Decision pro-cesses in verifying category membership statements:Implications for models of semantic memory. CognitivePsychology, 11, 1-37.

Meyer, D. E. (1970). On the representation and retrievalof stored semantic information. Cognitive Psychology,21, 242-300.

Miller, O. A. (1951). Language and communication. NewYork: McGraw-Hill.

Miller, G. A., & Johnson-Laird, P. N. (1976). Languageand perception. Cambridge, MA: Harvard UniversityPress.

Montague, R. (1974), Formal philosophy: Selected papers.New Haven, CT: Yale University Press.

Moran, L. (1941). Generality of word association responsesets. Psychological Monographs, 30, (4, Whole No.612).

Norman, D. A., Rumelhart, D. E., & the LNR ResearchGroup (Eds.). (1975). Explorations in cognition. SanFrancisco: Freeman.

Nunberg, G. (1978), The pragmatics of reference. Bloom-ington, IN: Indiana University Linguistics Club.

Perfetti, C. A. (1967). A study of denotative similaritywith restricted word associations. Journal of VerbalLearning and Verbal Behavior, 6, 788-795.

Putnam, H. (1975). The meaning of "meaning." In K.Gunderson (Ed.), Language, mind and knowledge:Minnesota studies in the philosophy of science (Vol. 7,pp. 131-193). Minneapolis: University of MinnesotaPress.

Quillian, M. R. (1968). Semantic memory. In M. L.Minsky (Ed.), Semantic information processing (pp.227-259). Cambridge, MA: MIT Press.

Raphael, B. (1968). SIR; A computer program for se-mantic information retrieval. In M. L. Minsky (Ed,),Semantic information processing (pp. 33-110). Cam-bridge, MA: MIT Press.

Riegel, K. F., & Riegel, R. M. (1963). An investigationinto denotative aspects of word meaning. Languageand Speech, 6, 5-21.

Rips, L. J., Shoben, E. J., & Smith, E. E. (1973).Semantic distance and the verification of semanticrelations. Journal of Verbal Learning and Verbal Be-havior, 12, 1-20.

Rogers, H. (1967). Theory of recursive functions andeffective computability. New York: McGraw-Hill.

Rosch, E. (1973). On the internal structure of perceptualand semantic categories. In T. M. Moore (Ed.), Cog-nitive development and the acquisition of language (pp.111-144). New York: Academic Press.

Rosch, E. (1976). Classification of real-world objects:Origins and representations in cognition. In S. Ehrlich& E. Tulving (Eds.), La memoire semantique (pp.242-250). Paris: Bulletin de Psychologic. (Reprinted inP. N. Johnson-Laird & P. C, Wason (Eds.), Thinking:Readings in cognitive science (pp. 212-222). Cam-bridge, England: Cambridge University Press, 1976).

Rumelhart, D, E,, & Levin, J. A. (1975). A languagecomprehension system. In D. A. Norman, D. E.Rumelhart, & the LNR Research Group (Eds.), Ex-ploration in cognition (pp. 179-208). San Francisco:Freeman.

Rumelhart, D. E., Lindsay, P. H., & Norman, D. A.(1972). A process model for long-term memory. In E.Tulving & W. Donaldson (Eds.), Organization andmemory (pp. 198-246). New York: Academic Press.

Schaeffer, B., & Wallace, R. (1970). The comparison ofword meanings. Journal of Experimental Psychology,86, 144-152.

Shoben, E. J., Wescourt, K. T., & Smith, E. E. (1978).Sentence verification, sentence recognition, and thesemantic-episodic distinction. Journal of ExperimentalPsychology: Human Learning and Memory, 4, 304-317.

Smith, E. E., Shoben, E. J., & Rips, L. J. (1974).Structure and process in semantic memory: A featuralmodel for semantic decisions. Psychological Review,81, 214-241.

Steinberg, D. D. (1970). Analyticity, amphigory, and thesemantic interpretation of sentences. Journal of VerbalLearning and Verbal Behavior, 1, 37-51.

Tabossi, P., & Johnson-Laird, P. N. (1980). Linguisticcontext and the priming of semantic information.Quarterly Journal of Experimental Psychology, 32,595-603.

Walpole, H. R. (1941). Semantics: The nature of wordsand their meanings. New Vbrk: Norton,

Warren, H. C. (1921). A history of association in psy-chology. New York: Scribner's.

Weinreich, U. (1966). Explorations in semantic theory.In T. A. Sebeok (Ed.), Current trends in linguistics,Vol. 3 (pp. 395-477). The Hague, The Netherlands:Mouton.

Woods, W. A. (1975). What's in a link: Foundations forsemantic networks. In D. G, Bobrow & A. M. Collins(Eds.), Representation and understanding: Studies incognitive science (pp. 35-82). New York: AcademicPress..

Received September 9, 1983Revision received April 5, 1984

Documents

Only Connections: A Critique of Semantic Networks