The Cambridge Handbook of Psycholinguistics
Edited by MICHAEL J. SPI,TEY University of California, Merced KEN MCRAE University of J17estern Ontario MARC F. JOANISSE University of J17estern Ontario
... ,.~.,... CAMBRIDGE ::: UNIVERSITY PRESS CHAPTER 12
The Human Conceptual System
Lawrence W. Barsalou
The human conceptual system contains i Recording versus interpretative people's knowledge about the world. systems Rather than containing holistic images of experience, the conceptual system repre The distinction between a recording sys sents components of experience, including tem and an interpretive system is central knowledge about settings, objects, people, to characterizing conceptual systems (e.g., actions, events, mental states, properties, Barsalou, 1999b; Dretske, 1995; Haugeland, and relations. Componential knowledge in 1991; Pylyshyn, 1973). A recording system the conceptual system supports a wide vari captures information about a situation by ety of basic cognitive operations, including creating attenuated (not exact) copies of it. categorization, inference, the representation Cameras, video recorders, and audio record of propositions, and the productive creation ers constitute good examples of recording of novel conceptualizations. In turn, these systems, each capturing records of experi basic operations support the spectrum of ence (e.g., photos, videos, audiotapes). A complex cognitive activities, including high recording system does not interpret ·what level perception, attention, memory, lan each component of a recording contains - it guage, thought, and socio-cultural cognition. simply creates an attenuated copy. For exam Traditional theories of Good-Old-Fashioned ple, a photo of a wedding records the light Artificial Intelligence (GOFAI), such as present at each point in the scene without semantic memory, constitute the dominant interpreting the types of entities and events approach to the conceptual system. More present. recently, researchers have developed alter Conversely, a conceptual system inter native approaches, including connectionist prets the entities perceived in an experience theories and simulation/embodied/situated or in a recording of one. To interpret a wed theories. ding, the human conceptual system might BARSALOU construe perceived individuals as instances somewhat like a recording, based both of bride, chair, cake, and so forth.' To achieve on experiential qualities and also on the interpretation, the conceptual system binds numerous feature areas in the brain that specific individuals in perception to knowl are mapped topographically, tonotopical!y, edge about components of experience in and somatotopically (e.g., Bear, Connors, memory. This is essentially the process of and Paradiso, 2001). Although imagery and categorization. A system that only records perception depart significantly from record perceptual experience does not categorize ings in important ways (e.g., Chambers and individuals in this manner. Instead, it simply Reisberg, i992; Hochberg, i998), they nev records them in the holistic context of an ertheless appear to have image-like quali undifferentiated sc€ne. ties such as orientation, extent, resolution, Interpretation supports other powerful vividness, and so forth. Thus, the argument computational abilities besides categoriza here is not that the brain lacks anything tion. Interpretation supports the produc like recording systems. To the extent that tion of inferences, allowing the cognitive the brain represents images in perception system to go beyond perceptual input. and imagery, it appears to utilize recording Interpretation supports the formulation of like representations. Instead, the argument propositions, where a proposition is a repre is that the brain also contains conceptual sentational structure that binds a concept representations used to interpret image (type) to an individual (token) in a manner like representations, thereby implementing that is true or false. Interpretation is produc powerful computational functions such as tive, supporting the construction of complex categorization, inference, propositions, and conceptual representations from simpler productivity. ones. Because the conceptual system sup Selective attention and memory integra ports these basic functions, it provides the tion are central to creating the conceptual larger cognitive system with computational knowledge that underlies interpretive pro abilities not possible in recording systems. cessing (Barsalou, i999b; 2003a). Whenever Cameras and other recording devices have selective attention focuses consistently on limited, if any, ability to implement cat some component of experience, conceptual egorization, inference, propositions, and knowledge about the component develops productivity. (cf. Schyns, Goldstone, and Thibaut, i998). Each time the component is attended, the information extracted becomes integrated i.1 Perceptual versus conceptual with past information about the same com representations ponent in memory. vVhen attention focuses Because recent theories propose that cate on a green patch of color, for example, the gory knowledge is grounded in the brain's information extracted is stored with previ modality-specific systems, it is useful to ous memories of green, thereby establishing establish a distinction between representa conceptual knowledge for this component. tions that are perceptual versus those that Over time, myriad components ofexperience are conceptual. Much work suggests that accumulate memories in a similar manner, the brain produces mental images that are including objects, events, locations, times, much like recordings (e.g., Kosslyn, i980; introspective states, relations, roles, proper i994). Furthermore, perceptual experi ties, and so forth. As conceptual knowledge ence can also be viewed as being at least abo~t these components develops, it can be used to interpret regions of perception and imagery, as described in greater detail later. i Italics \Vill be used to indicate concepts, and quotes Thus, perceptual and conceptual represen \Vill be used to indicate linguistic fqrrns ('i.vords, sen tences). Thus, bride indicates a concept, and "bride" tations work together to achieve cognitive indicates the corresponding \Vord. processing. THE HUMAN CONCEPTUAL SYSTEM
:h z Basic operations in a conceptual produces naming. On perceiving a robin, for te system example, conceptual knowledge for robin 1t becomes active to categorize it. In turn, the y, Once a system of conceptual knowledge word "robin" becomes active to name both ·s, develops for components of experience, the perceived individual and the conceptual .d it supports basic conceptual operations, knowledge activated, where the actual word l which in turn support more complex cogni produced is an individual instance of the d tive activities. As just described, these basic word category. Even when naming is implicit ,_ operations include categorization, infer (i.e., subvocal), this can be viewed as the i- ence, propositions, and productivity. Each is production of a word instance, grounded in 1, described in further detail here. Their roles a motor and auditory simulation. lt in complex cognitive activities are addressed Finally, recent work suggests that men .g later. tal simulations are central to linguistic pro lt cessing (e.g., Glenberg et al., 2005; Spivey, n Richardson, and Gonzalez-Marquez, 2005; 2.i Categorization ,_ Zwaan and Madden, 2005). To the extent ' lt During the process of categorization, the that meaning is represented this way, cate 11 cognitive system assigns perceived individu gorizing components of mental simulations ,_ als in perception and imagery to units of is central to linguistic processing. For exam g conceptual knowledge. While perceiving a ple, examining a simulation and categorizing IS soccer match, for example, individual set its components would be central to the pro cl tings (field), people (goalie), object5 (ball), cess of language production. Categorizing actions (kick), mental states (elation), and the components of simulation activates
1- so forth are assigned to categories. vVhile associated words, which are produced tJ imagining a soccer match, imagined indi in utterances to describe the simulation. ,_ viduals in the simulated perception can be Analogously, categorizing components of a ·r categorized similarly. perceived scene similarly underlies the pro n Categorization not only occurs in vision duction of an utterance to describe an actual 11 but in all modalities of experience. Thus, perception. In addition, categorizing regions s auditory events can be categorized (beep), as of a simulated or perceived scene not men ). can actions (walk), tactile sensations (soft), tioned explicitly produces inferences (e.g., e tastes (sweet), smells (pungent), affect inferring knife from an unlabeled region of :I (boredom), motivation (hunger), cognitive the simulation produced by the sentence, states (disbelief), and cognitive operations "Jeffrey cut the sandwich in half"). s (comparison). Furthermore, categorization e is central to processing all units of linguis 2.2 Inference tic analysis, including phonemes ("ba"), ver g balized words ("hello"), and written words An important theme in categorization ("exit"). In each case, a linguistic entity is research is that categorization is not an end categorized as an instance of a phoneme or itself (e.g., Markman and Ross, 2003). Simply r, word. Categorization is similarly central to knowing the category to which a perceived ;, identifying syntactic units (noun phrase) individual belongs is not particularly useful. and speech acts (question). Thus, categori- · What is useful are the inferential capabili zation is not only central to processing the ties that result. meaning of language but also to processing Once an individual has been assigned l its structure. correctly to a category, a multitude of use The semantic and structural aspects of ful inferences follow from associated con language are aligned (Langacker, i986). For ceptual knowledge that go beyond what example, categorizing nonlinguistic aspects has been perceived thus far for the indi of the world typically (but not always) vidual. Imagine perceiving and categorizing BARSALOU an unfamiliar individual as a cat. Useful system plays a central role in constructing inferences about the individual's structure, the meaning of a text. behavior, and internal states include that The conceptualizations that underlie lan the cat has teeth and claws, that it can purr guage production are similarly assumed to and scratch, and that it could be hungry and rely on systems of propositions. As people grateful. Useful inferences about relevant conceptualize what they want to describe, actions that the perceiver could perform fol they categorize individuals related to the low as well; such as being cautious toward topic under discussion, which produces the cat, petting it, and feeding it. Many type-token propositions (e.g., Bock, 1987). other potentially useful inferences also fol In turn, larger propositions, constructed low, including that the cat had a mother from conceptual predicates, result from and father (potentially relevant for breeding combining simpler ones. As the proposi purposes), that it could carry disease (rele tional representation develops, concepts vant for health purposes), and so on. Once in it activate associated words and syntac integrated conceptual knowledge about cat tic structures, which then surface in utter becomes active during categorization, a vari ances. The conceptual system provides a ety of associated inferences follow. fundamental link between the specific sit uation being described and the words used to describe it. 2. 3 Propositions Theories of psycholinguistics typically 2.4 Productivity assume that propositional representations underlie the meanings of comprehended The human cognitive system can produce texts (e.g., Kintsch and van Dijk, 1978). an infinite number of linguistic and con Most simply, a proposition can be viewed as ceptual structures that go far beyond those a type-token relation that becomes estab experienced. No one ever experienced a real lished between an individual and a concept. Cheshire cat, but it is easy to imagine and Thus, the process of categorization described then describe "a cat whose body fades and earlier produces propositions. Categorizing reappears while its human smile remains." an individual chicken, for example, creates Similarly, it is possible to begin with the a proposition that consists of the individual conceptualization of a familiar object and chicken (a token) being bound to the con then to imagine it in nonexperienced forms, cept for chicken (a type). In text comprehen such as conceptualizing a gray cat and then sion, similar type-token propositions arise conceptualizing it as a purple cat or as a pur as the meanings of words are combined. ple cat with green polka dots. Hearing "Ralph is a chicken," for example, Productivity underlies people's creative produces the proposition, chicken (Ralph), abilities to combine words and concepts where the notation used is type (token). As into complex linguistic and conceptual this example illustrates, chicken is a predi structures compositionally (e.g., Fodor and cate that takes individuals as arguments, Pylyshyn, 1988; also see Barsalou, 1999b; such as Ralph. Other concepts take multi 2003a). Productivity generally appears to ple arguments, in particular, verbs and prep result from combinatorial and recursive ositions. For example, the verb eat can take mechanisms. Combinatorial mechanisms arguments for agent, patient, and instrument, allow people to take a word (or concept), as in eat (John, soup, spoon). While compre and then rotate other words (or concepts) hending phrases, sentences, and texts, many through a particular relation associated with elemental propositions like these are con it. Beginning with the noun "cat," for exam· structed, which are then assembled hier pie, noun phrases ~an be. constructed com· archically into larger and more complex binatorially by rotating other words through propositional structures. Because the types a modifier relation, thereby creating "gray in propositions are concepts, the conceptual cat, II I/ orange cat, II ll purp1 e cat, ll H pm' k ca t ,» THE HU1'1AN CONCEPTUAL SYSTEM 243
and so forth. Similarly, nouns can be com cognitive processes such as perception and binatorially rotated through the thematic attention. As we will see, however, concep 1- roles associated a particular verb, such as tual knowledge permeates every aspect of :o rotating" cake," 11 pizza/1 and "tamale" through cognition from high to low. vVithout knowl le the patient role of"eat" (other nouns could edge, any cognitive process would stum e, similarly be rotated through other roles for ble into ineffectiveness. There is no such 11 e 11 eat," such as ufork" and fingers," for the thing as a knowledge-free cognitive process. 's instrument): To understand cognition, it is essential to ). In recursion, complex conceptual and understand the conceptual system and its ::I linguistic structures are nested within exist ubiquitous presence across the spectrum of 1 ing linguistic and conceptual structures. cognitive activities. When conceptualizing a face, for example, s people could first conceptualize a head. 3.1 High-level perception Nested within the conceptualization of the head, people could then conceptualize the As people interact with the environment eyes, then the eyeballs, then the irises, and and attempt to achieve goals, the conceptual so forth. Analogously, people can describe system supports the construction of percep this embedded conceptual structure lin tions. For example, conceptual knowledge guistically, as in "the head contains the eyes, contributes to the mechanisms that sepa which contain eyeballs, which contain irises, rate figure from ground (e.g., Peterson and and so forth." Embedding conceptual and Gibson, 1994), and also to processes that fill linguistic structures within other structures in missing regions of incomplete percep allows people to construct novel conceptu tual experiences (e.g., Palmer, 1999; Samuel, alizations and verbalizations not encoun 1997). Conceptual knowledge produces tered previously. anticipation inferences about what is likely In summary, using combinatoric and to happen next (e.g., Reed and Vinson, recursive mechanisms, people construct an 1996), and also the specific forms that these unlimited number of complex represen anticipations take (e.g., Shiffrar and Freyd, tations from finite numbers of words and 1993; Stevens et al., 2000). Finally, concep concepts. This ability appears to result from tual knowledge helps to predict entities and a productive system for language that is events likely to be present in the current closely coupled to a productive system for scene, thereby speeding their categorization conceptualization. It is generally assumed (e.g., Biederman, 1981; Palmer, 1975; Yeh and that these two systems have parallel struc Barsalou, 2006). ture (e.g., Langacker, 1986). As a result, constructing linguistic expressions produc 3.2 Selective attention tively produces corresponding conceptual structures. Conversely, constructing con Once a concept becomes active to construe ceptualizations productively produces cor a situation, it controls the distribution of responding linguistic descriptions. attention across it; For example, when the concept for a spatial preposition becomes active (e.g., above), it directs attention to a 3 The conceptual system supports likely region where a focal figure will appear the spectrum of complex cognitive relative to the ground below. Specifically, the activities ideal position is for the figure to be aligned geometrically above the center ofthe ground, Researchers often assume that the concep not too far away. On hearing "the square is tual system resides in the province of higher above the circle," for example, people gen cognition along with language and thought. erally infer that the square is center aligned Conversely, researchers often assume that above the circle, not too far away. Much the conceptual system is irrelevant to lower work demonstrates that spatial concepts T I ! BARSALOU direct attention to prototypical locations in Carmichael, Hogan, and vValter, i932; Craik this manner (e.g., Carlson-Radvansky and and Lockhart, i972; Huttenlocher, Hedges, Logan, i997; Hayward and Tarr, i995; Logan and Duncan, i991). Rather than solely cap and Compton, i996). After reading the word turing perceptual images as does a camera for a spatial location, the activated spatial or video recorder, the brain encodes images concept directs attention to the most likely together with concepts that interpret them. position in the display. As a result, the memory of a stimulus con Additional research shows that infer tains both perceptual and conceptual infor ences about function modify these atten mation. Once a stimulus is encoded, it tional inferences (e.g., Carlson-Radvansky, becomes stored together with other mem Covey, and Lattanzi, i999; Conventry, ories encoded previously with similar con i998). Consider the statement "the tooth ceptual structures. Much work shows that paste tube is above the toothbrush." If spa as the number of memories stored with a tial geometry were the only factor affecting concept increases (i.e., fan), interference attentional inferences, then a picture of between the memories becomes more a toothpaste tube centered geometrically severe (e.g., Anderson, i976; Postman and over a toothbrush should be verified faster Underwood, i973). Finally, concepts further than when the two objects are not centered become active during memory retrieval to geometrically. Verification is fastest, how produce classic reconstruction effects (e.g., ever, when the toothpaste tube is positioned Bartlett, i932; Brewer and Treyens, i981). functionally (not geometrically) over the Thus, concepts enter ubiquitously into all end of the toothbrush having the bristles. phases of memory processing. Thus, the concept above does not trigger a single attentional inference based on ideal 3.4 Language ized geometry. Instead, the noun concepts combined with above during the construc The semantics of natural language are tion of propositions jointly determine the closely related to the human conceptual inference. system. Although lexical meanings are not identical to concepts, the two have much in common and influence each other exten 3. 3 Episodic memory sively (e.g., Barsalou et al., i993; Marslen Besides being central to online process Wilson, i992; Schwanenflugel, i991). The ing of the environment, the conceptual access of word meaning can be viewed as system is central to offline processing in an inferential process. On perceiving a word memory, language, and thought. In each of such as "bird," retrieving semantic informa these complex cognitive activities, process tion constitutes inferences about the word's ing a nonpresent situation is often of pri meaning. American readers are more likely, mary importance, with perception of the for example, to infer that "bird" means current environment being suppressed to something having the properties of small, facilitate processing the imagined situation flies, and sings, rather than something having (Glenberg, Schroeder, and Robertson, i998). the properties of large, runs, and squawks. Humans are much more adept at represent Typically, these meanings are highly context ing non present situations than other species; dependent, reflecting both the surrounding with the control of conceptual representa text and the pragmatics of the communica tions via language appearing central to this tive sii:uation (e.g., Barsalou, i999a; Yeh and ability (e.g., Donald, i993). Barsalou, 2006). The conceptual system enters into all As the meanings of words become com three classic phases of memory activity: bined during the construction of proposi encoding, storage, and retrieval. During tions, background conceptual knowledge is encoding, the conceptual system pro used extensively. In particular, knowledge vides diverse forms of elaboration (e.g., of conceptual relations is often central to THE HUMAN CONCEPTUAL SYSTEM 245
integrating word meanings (e.g., Gagne and of background knowledge are retrieved Shoben, 1997; 'Wisniewski, 1997). For exam and incorporated into the decision making ple, integrating the mea_nings of lake and process. Loken, Barsalou, and Joiner (2008) trout to understand "lake trout" requires document a wide variety of roles that con activating knowledge about the relation ceptual processes play in consumer decision LOCATION (X, Y), whereas integrating making. the meanings of swinging and vine to under The conceptual system is also central to stand "swinging vine" requires activating reasoning. vVhile performing deductive rea knowledge about the relation MOTION soning, people do not simply manipulate (X, Y). abstract logical expressions. Instead, they Inference production beyond individual appear to manipulate conceptual represen words is a well-established aspect of lan tations about the reasoning domain, thereby guage comprehension (e.g., Bransford and exhibiting widespread content effects (e.g., Johnson, 1973; Schank and Abelson, 1977). Cheng and Holyoak, 1985; Johnson-Laird, As people comprehend a text, they infer 1983). Conceptual representations are also considerable amounts of background knowl central to inductive reasoning, especially edge not stated explicitly. For example, when it concerns categories (e.g., Medin comprehenders infer a variety of thematic et al., 2003). Finally, conceptual representa roles, such as hearing "Mary pounded a nail tions are central to causal reasoning across into the wall" and inferring that a hammer a wide variety of domains, including clini was used (e.g., McRae, Spivey-Knowlton, cal diagnosis (e.g., Kim and Ahn, 2002) and and Tanenhaus, 1998). Similarly, when peo artifact function (e.g., Barsalou, Sloman, and ple hear the sentence "The surgeon put Chaigneau, 2005). on gloves before beginning the operation," Problem solving also relies extensively they are surprised when the next sentence on conceptual processes. Similar to reason begins "She was tired from the previous ing, widespread effects of domain-specific operation," because they make default gen knowledge occur (e.g., Newell and Simon, der inferences (e.g., Carreiras et al., 1996). 1972). The same abstract problem can be In general, the more deeply people compre difficult to solve when grounded in one hend a text, the richer the inferences they domain but easy when grounded in another, produce, not only about thematic roles but depending on the availability of relevant about explanations and a wide variety of knowledge. Ross (1996) argues further that other conceptual structures (e.g., Graesser, knowing how to use artifacts for solving Singer, and Trabasso, 1994). Researchers typ problems constitutes a significant aspect ically assume that these rich comprehension of category knowledge. Rather than simply inferences arise via the conceptual system containing physical features that identify as relevant conceptual knowledge becomes category members, a category representa active. tion contains extensive knowledge about how to use its exemplars for achieving goals (also see Barsalou, 1991). 3. 5 Thought Thought requires extensive use of concep 3.6 Social and cultural cognition tual representations. As people perform decision making, reasoning, and problem The conceptual system plays extensive roles solving, conceptual representations become in sociai cognition (e.g., Fiske and Taylor, activated as the objects of thought. During 1991; Kunda, 1999). During social interac decision making, the choice objects under tion, people use socia.l knowledge to catego consideration are represented conceptu rize perceived individuals into social groups. ally (e.g., Markman and Medin, 2002). As Stereotypes for these groups can be viewed possible choice objects are evaluated, fea as conceptual representations that have been tures, relations, values, and diverse forms distorted by various sources of background BARSALOU knowledge. Once a perceived individual has wider variety is possible. A relatively generic been assigned to a social category, rich infer description of each approach will serve to ences (attributions) result about the causes illustrate it. of the person's behavior, their mental state, and likely actions. Self-concepts constitute 4.1 GOFAI theories another central form of conceptual knowl edge in the social domain. GO FAI theories of the conceptual system Although the basis of a culture can be originated in artificial intelligence during localized in its artifacts, activities, organi the cognitive revolution (e.g., Haugeland, zations, and institutions to a considerable i985). To represent knowledge in comput extent, it can also be localized in concep ers, artificial intelligence researchers devel tual knowledge of these external entities oped new representation languages based (e.g., Shore, i996). Cultural transmission on predicate calculus (e.g., Charniak and can be viewed, in part, as the propagation McDermott, i985; Newell and Simon, i972). of conceptual knowledge from generation Typically, these representation languages to generation, along with the transmission included predicates to represent conceptual of other things, such as skills. Much recent relations, arguments that become bound to work illustrates that different conceptual values, and recursive nesting that embeds knowledge produces major cognitive and predicates within predicates (e.g., Barsalou, behavioral differences among cultures (e.g., i992). Reflecting the goals of knowledge Atran, Medin, and Ross, 2005). engineering, the GO FAI representation of a concept typically contains an extensive amount of information, such that a given 4 Theories of the conceptual system concept contains many propositions, If a computer is to have sufficient knowledge Three approaches to theorizing about the for understanding language, answering ques conceptual system enjoy varying degrees of tions, and solving problems, its knowledge acceptance in psychology, cognitive science, must be extensive. and cognitive neuroscience. The most tra In contrast, psychological versions of ditional theories, and perhaps still the most GO FAI theories are typically much sparser, dominant, originated in what Haugeland reflecting the goal of testing psychological (1985) dubbed "GOFAI" for Good Old models in a controlled and rigorous man fashioned Artificial Intelligence. In partic ner. Thus, psychological versions likely con ular, the theory of semantic memory con siderably underestimate the complexity of stitutes perhaps the best known and most naturally occurring conceptual representa widely accepted view of the conceptual sys tions (e.g., Barsalou and Hale, i993). Two tem. Connectionist theories constitute a sec general subclasses of the GOFAI approach ond major class of theories. This approach have dominated theories of the concep reflects an increasing appreciation of neu tual system and continue to do so: semantic ral mechanisms and statistical processing, memory and exemplar models. The seman both relatively absent in GOFAI theories. tic memory view, in particular, continues to Simulation, embodied, and situated theo constitute the primary way that research ries constitute the most recent class. ·while ers in many communities think about the incorporating neural and statistical mecha conceptual system. Researchers across psy nisms, they further emphasize the brain's chology, ,cognitive science, and cognitive modality-specific systems, the body, and the neuroscience implicitly adopt the semantic environment. memory framework when they must address Each of these three approaches is knowledge in their respective research areas. described next. Within each approach, a Semantic memory and exemplar models are wide variety of models exists, and an even each addressed in turn. THE HUl;lAN CONCEPTUAL SYSTEM 247
in the modalities and for the environmental 1 SEMANTIC MEMORY f~~ construct of semantic memory arose entities they represent. from a proposed distinction between seman Representations in semantic memory tic and episodic memory (Tulving, 1972). are also generally assumed to be relatively Specific examples in~l~de the network mod abstract and decontextualized. In the typi els of Collins and Quillian (1969), Collms and cal theory, the representation of a category Loftus (1975), and Glass and Holyoak (1975). is a prototype or rule that distills relatively As Hollan (1975) notes, prototype and other invariant properties from exemplars. Lost in feature set models (e.g., Reed, 1972; Rosch the distillation are idiosyncratic properties of and Mervis, 1975) are roughly equivalent to exemplars and background situations. Thus their network counterparts, together form the representation of chair might be a decon >ed ing a more general class of semantic mem textualized prototype that includes seat, back, nd ory models. Thus, semantic network, feature and legs, with idiosyncratic properties and '2). list, and prototype models will be subsumed background situations filtered out. Although ~es here under the larger rubric of semantic functional properties may be extracted and ial memory. For further review of these models, stored, they typically tend to be decontex to see Smith (i978). tualized invariants, not detailed informa els Following Tulving's classic proposal, tion about specific situations. The resulting semantic memory is widely viewed as mod- representations have the flavor ofdetached . · ular, that is, as an autonomous system sepa encyclopedia descriptions in a database of rate from the episodic memory system. Less categorical knowledge about the world. explicitly, but equally true, semantic mem Similar to being decontextualized, n ory is also viewed widely as separate from semantic memory representations are typ a the brain's modality-specific systems. It is ically viewed as relatively stable. For a given e generally assumed that semantic memory · category, these theories assume that differ does not share representation and process ent people share roughly the same repre ing mechanisms with perception, action, sentation, and that the same person uses the and interoception,' but is instead a relatively same representation on different occasions. f independent system with its own principles Finally, semantic memory models excel of representation and processing. in implementing the basic operations of One of these distinguishing principles propositions and productivity described is representational format, namely, repre earlier. Because the representations in these sentations in semantic memory are widely models typically include predicates whose viewed as amodal. Rather than being rep arguments become bound to values, with resentations in modality-specific systems, the potential for predicates to embed recur semantic memory representations are typ sively, they naturally implement proposi ically viewed as redescriptions of modality tions and productivity. Although semantic specific states in an amodal representation memory models can implement categori language, namely, one that lacks modality zation and inference using prototypes· and specific qualities. For example, the concep definitions, they have been widely criticized tual representation of the visual property as being too abstract and rigid in how they red is an amodal symbol that stands for perform these basic operations. Typically, perceptual states of red in the visual sys . semantic memory models are not sensitive I tem and their physical counterparts in the to the details of exemplars and situations world. In general, amodal representations in and do not contain adaptive mechanisms semantic memory stand for representations that implement learning.
4-L2 EXEMPLAR MODELS 2 Interoception here \Vill refer to the perception of Since Medin and Schaffer's (1978) context internal states, namely1 states of motivation, emotion, and cognition that are accessible to consciousness. model, exemplar models have provided -'Y''-,<:1.. ·· ' BARSALOU a strong competitor to semantic memory to assume that all exemplar memories for models. Many important variants of the a category are accessed every time the cate basic exemplar model have been developed, gory is processed. Although an exemplar set including Nosofsky (1984), Heit (1998), and can be very large, its constant application Lamberts (1998). Exemplar models are across different occasions is relatively stable, included within the broader class of GOFAI with all exemplars being applied. Exemplar models because they tend to use standard models that sample small subsets of exem symbolic notation for expressing the prop plars, on the other hand, are dynamic (e.g., erties of exemplars, unlike connectionist Barsalou, Huttenlocher, and Lamberts, 1999; theories and simulation/embodied/situated Nosofsky and Palmeri, 1997). theories, which use statistical and neural ·where exemplar models excel is in cat representation languages. egorization. Because extensive detail about Architecturally, exemplar models tend to a category is stored - both in terms of idi be modular in that exemplar knowledge is osyncratic exemplar properties and back again assumed implicitly to reside in mem ground situations - these models are highly ory stores outside the brain's modality-spe accurate during categorization and can cific systems. Similar to semantic memory adapt quickly to changing category infor models, redescriptions in an amodal rep mation. Although exemplar models have resentation language typically capture the not been developed to explain inference, content of exemplar memories, standing in they can in principle produce highly accu for the modality-specific states experienced rate inferences following categorization, originally. again because of the large amounts of infor Notably, however, some exemplar mation stored and the context-specificity models view exemplar representations as of retrieval processes that operate. on it. implicit memories in modality-specific sys ·where exemplar models are weakest is on tems (e.g., Brooks, 1978; Jacoby and Brooks, symbolic operations. Thus far, this approach 1984; cf. Roediger and McDermott, 1993). has not attempted to implement predicates, According to this approach, for example, arguments, and recursion, and therefore an exemplar for a visual category is stored does not implement the basic operations of as a visual memory in the visual system, propositions and productivity. not as an amodal description outside it. Exemplar models that store exemplars in Connectionist theories modality-specific systems can be construed +2 as nonmodular, given that common repre Feedforward connectionist networks con sentations underlie both conceptual and stitute a relatively recent but increasingly modality-specific processing. influential theory of the conceptual system. ·where exemplar models differ most from For general accounts of feedforward nets, see semantic memory models is on abstraction Rumelhart, Hinton, and vVilliams (1986) and and decontextualization. ·whereas seman Bechtel and Abrahamsen (2002). For specific tic memory models distill properties across applications of the foedforward architecture exemplars and store them as abstractions to representing conceptual knowledge, see (e.g., prototypes and rules), exemplar mod Hinton (1989), Kruschke (1992), Rumelhart els simply store exemplar memories, thereby and Todd (1993), Tyler et al., (2000), and capturing idiosyncratic information about Rogers and McClelland (2004). A variety of category instances along with details about other c6nnectionist architectures have also the situations in which they occur. been used to model the conceptual system, Perhaps counterintuitively, exemplar which are not addressed here (e.g., Cree, models tend to assume that category rep McRae, and McNorgan, 1999; Farah and resentations are relatively stable, much like McClelland, 1991; Humphreys and Forde, semantic memory models. Stability exists in 2001; McClelland and Rumelhart, 1985; most exemplar models because they tend Rumelhart et al., 1986). THE HU~IAN CONCEPTUAL SYSTE~I 249 s for Perhaps surprisingly, feedforward nets, modality-specific and conceptual represen cate like GOFAI theories, implement a modu tations reside in different modular systems, Lr set lar conceptual system. Whereas the input with arbitrary mappings between them. ition layer of a feedforward net is interpreted No doubt, other significant aspects of the able, as a perceptual system, its hidden layer is representations differ, with connectionist tplar viewed as 'implementing conceptual rep representations being statistical, and seman :em resentations. Thus one "module" of units tic memory representations being discrete. e.g., underlies perception, and a second module Nevertheless both approaches contain 999; underlies conception, thereby establishing a amodal redescriptions of perceptual input at modular distinction between them. Because a general level of analysis. cat- complex interactions can arise between vVhere feedforward nets depart most 1out these two systems, they are not modu notably from semantic memory models is on idi lar in the sense of being impenetrable (cf abstraction and stability (similar to exemplar LCk Fodor, i983; Pylyshyn, i984). Nevertheless models). Rather than establishing decontex hly different representational systems underlie tualized representations of categories, feed can perception and cognition, such that mod forward nets store situated representations :or-c ularity exists in a somewhat nonstandard in two ways. First, these nets acquire much ave sense. As will be seen shortly, it is possible idiosyncratic information about exemplars lCe, to formulate a conceptual system in which (as in exemplar models), rather than discard :u- shared neural units represent information in ing this information during the abstraction )Q, perception and conception. It is also worth of category invariants. Although invariants or- noting that some of the alternative connec may be abstracted implicitly, much idiosyn ity tionist architectures mentioned earlier oper cratic information is maintained that plays it. ate in this latter manner. Thus, modularity central roles in processing. Second, feedfor ::m only applies to connectionist nets that have ward nets store extensive information about ch feedforward architectures, along with other the situations in which exemplars occur. architectures that use separate pools of units Rather than extracting focal knowledge of re for perception and conception. a particular category instance from a back of Because of this modular architecture, ground situation, much correlated informa internal representations in feedforward nets tion about the situation is stored as well (e.g., are amodal. Before learning begins, connec Rumelhart et al., i986). As a consequence, tions between the input and hidden layers activating an exemplar typically retrieves are set initially to small random values so situational information and vice versa. l- that learning is possible. As a result, the par Feedforward nets are also highly dynamic. y ticular units in the hidden layer that become Rather than representing a category with a l. positively (or negatively) associated with stable representation, as in semantic mem e particular units in the input layer are deter ory and exemplar models, a feedforward net :! mined arbitrarily. The surprising implication uses a space of representations. Specifically, c is that statistical patterns on the hidden units a category's representation is an attractor ~ associated with particular categories func within the possible activation states of the tion as "fuzzy" amodal symbols, standing in hidden units, with an infinitely many states Il for their perceptual counterparts. With each around the attractor providing possible rep I new set of random starting weights, a dif resentations. On a given occasion, the repre ferent mapping develops.l The arbitrariness sentation activated to represent the category that results is much in the spirit of semantic is a function of the network's current state, i memory representations. In both approaches, input, and learning history. Thus a concept in a feedforward ne.t is a dynamic system that produces a family of representational j 3 It is worth noting that invariants exist across the l states, depending on current conditions. . I different mappings. Regardless, each mapping is a redescription of the input in a separate modular Like exemplar models, feedforward system. nets excel in categorization and inference. BARSALOU
Because extensive detail about a category cognitive linguistics adopts similar views is stored - both in terms of idiosyncratic (e.g., Fauconnier, 1985; Lakoff and Johnson, exemplar properties and background situ 1980, 1999; Langacker, 1986; Talmy, 1983), but ations - feedforward nets are highly accu have not yet typically drawn strong connec rate during categorization, and can adapt tions to cognitive and neural mechanisms quickly to changing category information. (although see Gallese and Lakoff, 2005). For the same reason, feedforward nets pro All of these approaches assume that the duce highly accurate inferences following conceptual system is nonmodular. Rather categorization. ·where connectionist models than having separate systems for modality are weakest (like exemplar models) is on specific and conceptual processing, a com symbolic operations (Fodor and Pylyshyn, mon representational system is assumed 1988). Although some attempts have been to underlie both. According to this view, made to implement predicates, arguments, conceptual processing relies heavily on and recursion (e.g., Pollack, 1990; Smolensky, modality-specific simulations to represent 1990), these approaches have not been categories (for more detail on the simula widely accepted as plausible psychological tion process, see Barsalou, 1999b; 2003a). or neural accounts of the conceptual system. A consequence of this nonmodular archi So far, connectionism has not succeeded in tecture is that conceptual representations convincing the cognitive psychology, cog are modal, not amodal. The same types of nitive science, and cognitive neuroscience representations underlie perception and communities that this approach explains conception. When the conceptual system the basic conceptual operations of proposi represents an object's visual properties, it tions and productivity. uses representations in the visual system; when it represents the actions performed on an object, it uses motor representations. 4. 3 Simulation, embodiment, and situated Depending on the distribution of modalities theories· on which people experience a category, a Recent theories have focused on the roles particular distribution of modality-specific of modality-specific simulation, embodi information becomes established for it (e.g., ment, and situations in conceptual process vision and taste for fruit versus vision and ing. Damasio (1989) 1 Martin (2001), Barsalou action for tools; Cree and McRae, 2003). (1999b; 2003a ), and Simmons and Barsalou Although perception and conception (2003) focus on modality-specific simulation. are similar in this framework, they are not Glenberg (1997) and Barsalou et al. (2003) identical. Whereas bottom-up mechanisms focus on embodiment. Barsalou (1999a; dominate the activation of modality-spe 2003b; 2005) and Barsalou, Niedenthal et al. cific systems during perception, top-down ( 2003) focus on situations. Although these mechanisms dominate during conception. approaches differ somewhat in emphasis, Furthermore, the representations acti they all assume that the conceptual sys vated in conception are partial .reenact tem specifically and cognition in general are ments of modality-specific states, and may grounded in the brain's modality-specific often exhibit bias and reconstructive error. systems, in the body, and in the environ Nevertheless, perception and conception ment. According to these approaches, the are far from being modular autonomous cognitive system is not self-sufficient but systems. depends in important ways on its ground The claim is not that modal reenactments ings. Indeed, these approaches assume that constitute the sole form of conceptual rep grounding mechanisms are central parts of resentation. As Simmons and Barsalou the cognitive system, not merely a periph (2003) suggest, representations in the brain's eral interface. For a recent collection of association areas also play a role, perhaps papers on this approach, see Pecher and somewhat analogous to the hidden unit Zwaan (2005). Much additional work in representations in connectionist nets. This is THE HUMAN CONCEPTUAL SYSTEM iews consistent with the widespread finding that (1991; 2003b ), ad hoc and goal-derived cat ison, other factors influence conceptual process egories develop to bind roles in action , but ing besides the modalities (e.g., statistical schemata with their instantiations in the nec strength, correlation, and uniqueness; Cree environment. As systems of these mappings isms and McRae, 2003; Tyler et al., 2000). Thus, develop, the conceptual system becomes the claim is simply that modal simula organized around the action-environment the tions are one important and widely utilized interface. th er form of -representation during conceptual lity- processing. 4· 3.1 COMMON MISCONCEPTIONS om Regarding abstraction and stability, this Three common misconceptions arise fre ned approach assumes that conceptual repre quentlyaboutsimulation/embodied/situated levv, sentations are dynamic and situated. Rather views. One is that they are purely empiricist on than being a single abstracted representation with no nativist contributions. Although ;ent for a category, a concept is a skill for con extreme empiricist views are possible and 1la- structing idiosyncratic representations tai sometimes taken, there is no a priori reason lored to the current needs of situated action why strong genetic constraints could not :hi- (Barsalou, 2003b). Actually, Barsalou, et al. underlie a system that relies heavily on sim ons (2003) advocate discarding the use of concept ulation, embodiment, and situatedness. For : of altogether and replacing it with accounts of example, specific simulations could in prin ind the specific mechanisms that represent cat ciple be determined genetically. More plau em egories. In this spirit, Barsalou (1999b; 2003a) sibly, however, strong genetic constraints ' it proposes the construct of a simulator as a may exist on the mechanisms that capture :m; distributed neural mechanism that con and implement simulations. In this spirit, 1ed structs an infinite set of specific simulations Simmons and Barsalou (2003) propose that ns. to represent a category, property, or relation the association and feature areas underlying ies dynamically. Thus, the simulator for chair simulations reflect constraints on categories ,a can construct many simulations of differ that developed over the course of evolution fie ent chairs, from different perspectives, used (also see Caramazza and Shelton, 1998). g., for different purposes, reflecting the agent's A second common misconception about nd current goal and situation. simulation/embodied/situated approaches A given simulation is assumed to repre is that they necessarily implement·record
)fl sent more than the focal category of ing systems and cannot implement concep ot interest. Additional information about tual systems for interpreting the world. As ns background settings, goal-directed actions, Barsalou (1999b; 2003a) proposes, however, e and introspective states is also assumed to modality-specific systems can implement rn be included, making simulations situated basic conceptual operations, such as cate n. (e.g., Barsalou, 1999a, 2003b, 2005; Barsalou, gorization, inference, propositions, and pro i Niedenthal, et al., 2003). On a given occa ductivity. The essential idea is that selective t- sion, a specific simulation is tailored to the attention extracts information about the 1y computational and pragmatic demands of components of experience to establish sim •f. the current situation. Thus, the conceptual ulators for these components. Once these n system is dynamic and situated, similar to simulators exist for object, events, mental IS feedforward nets, but with modal represen- · states, relations, properties, and so forth, the tations instead of amodal ones. argument is that they naturally implement :s A related theme is that the conceptual basic conceptual operations. system is organized around situated action A third common misconception is that l (cf Glenberg, 1997). A fundamental prob abstract concepts cannot be represented in s lem in situated action is mapping action simulation/embodied/situated approaches. s effectively into the world, and one possibil Various researchers, however, have argued t ity is that the conceptual system develops to that mechanisms within this approach are facilitate this process. According to Barsalou capable of representing these concepts. BARSALOU
For example, Lakoff and Johnson (1980; approach to cognitive linguistics rests on 1999) propose that abstract concepts are this assumption, namely, the linguistic sys grounded metaphorically in concrete con tem serves as an instrument for controlling cepts (but see Murphy, 1996 for a critique). the conceptual system. Alternatively, Barsalou (1999b) and Barsalou Increasing empirical research suggests and vVieiner-Hastings (2005) propose that that both the linguistic and conceptual sys abstract concepts are grounded in situated tems are active as people perform concep simulations, just like concrete concepts, but tual tasks (see Glaser, 1992 for a provocative focus on different situational content, espe review). Depending on task materials (e.g., cially on interoceptions and events. words versus pictures) and task conditions (e.g., superficial versus deep processing), 4-3·2 COMPUTATIONAL conceptual processing relies on varying mix IMPLEMENTATION tures of the linguistic and conceptual sys One major limitation of the simulation/ tems. Further evidence for this view comes embodied/situated approach to date is the from Solomon and Barsalou (2004) and Kan relative lack of computational frameworks et al., (2003). In these experiments, subjects for implementing it. Increasingly, however, used different mbctures of linguistic pro implementations are being developed. For cessing and simulation while verifying the example, Cangelosi and his colleagues have conceptual properties of objects under dif recently begun implementing the ground ferent task conditions. Barsalou et al., (2005) ing mechanisms in simulation/embodied/ offer further behavioral and neural evidence situated theories (e.g., Cangelosi, Greco, and that conceptual processing utilizes vary Hamad, 2000; Cangelosi et al., 2005; Cangelosi ing mixtures of linguistic processing and i; and Riga, 2005; Joyce et al., 2003). Also, the simulation. top-down mechanisms in O'Reilly's neural net architectures have significant poten tial for implementing simulations (e.g., 5 Conclusion O'Reilly, 1998, 2006). Other recent attempts to ground computational accounts of cogni · As reviewed here, three basic accounts of tion in modality-specific processing include the conceptual system exist in modern cog Roy (2005) and Clark and Mendez (2005). nitive psychology, cognitive science, and Acceptance of the simulation/embodied/sit cognitive neuroscience: ( 1) classic GO FAI uated approach clearly depends on increas approaches, such as semantic memory and ing formalization, but there appears to be exemplar models, that utilize amodal sym no a priori reason why formalization is not bols in a modular conceptual system; (2) sta possible. Given the relative recency of this tistical approaches, such as connectionism ! ' approach, together with the complexity of and neural nets, that implement dynamic the mechanisms that must be implemented, and situated conceptual representations; (3) it is not surprising that mature formal simulation/embodied/situated approaches accounts do not yet exist (for discussion of that ground conceptual knowledge in these complexities, see Barsalou, 1999b, pp. modality-specific systems, in the body, and 651-2). Of interest will be whether viable in the environment. computational accounts can be constructed Claiming that significant value exists in i '! in the coming years. all three approaches might seem unduly l i - diplomatic. To the contrary, however, each 4.3.3 RELATIONS BETWEEN LANGUAGE of these approaches has discovered some AND SIMULATION thing fundamentally important about the Finally, several lines of research propose human conceptual system. Classic GOFAI that the linguistic system is. closely coupled approaches have established the impor with the simulation system. As mentioned tance of propositional representations and earlier, a central tenet of Langacker's (1986) productivity in conceptual processing. THE HUMAN CONCEPTUAL SYSTEM 2 53 l Statistical approaches have highlighted the semantic organization (pp. 21-74). Hillsdale, importance of adaptation, generalization, NJ: Lawrence Erlbaum Associates. (1999a). Language comprehension: Archival 7 partial matching, frequency effects, and pat ' tern completion. Simulation/embodied/sit memory or preparation for situated action. Discourse Processes 28 61-80. uated approaches have drawn attention to 1 1 (1999b). Perceptual symbol systems. Behavioral the importance of grounding knowledge in and Brain Sciences, 22, 577-660. the brain's modality-specific systems, in the ( 2003a). Abstraction in perceptual symbol sys body, and in, the environment. tems. Philosophical Transactions of the Royal Barsalou ( i999b) ends with the following Society of London: Biological Sciences, 358, conjecture: Successful theories in the future u77-87. are likely to integrate all three frameworks ( 2003b). Situated simulation in the human into a single system (p. 652). It is unlikely conceptual system. Language and Cognitive that theories implementing only one or Processes, i8, 513-fo. . even two of these approaches will succeed. (2005). Continuity of the conceptual system ·what each approach offers appears essential across species. Trends in Cognitive Sciences, 9, to the human conceptual system. 301j-ll. Barsalou, L. W. & Hale, C.R. (1993). Components It is probably fair to say that GO FAI of conceptual representation: From feature and connectionist theories have gener lists to recursive frames. In Van Mechelen, I., ally attempted to incorporate only one, or Hampton, J., Michalski, R., & Theuns, P. (Eds.) occasionally two, of these approaches. In Categories and concepts: Theoretical views and contrast, simulation/embodied/situated inductive data analysis (pp. 97-1+1). San Diego, views have typically attempted to incorpo CA: Academic Press. rate two and sometimes three approaches, Barsalou, L. 'vV., Huttenlocher, J., & Lamberts, K. not only emphasizing grounding, but also (1998). Basing categorization on individuals emphasizing statistical processing and sym and events. Cognitive Psychology, 36, 203-72. bolic operations. Again, however, we have Barsalou, L. 'vV., Niedenthal, P. M., Barbey, A., yet to see· fully developed computational & Ruppert, J. (2003). Social embodiment. In Ross, B. (Ed.) The Psychology of Learning and accounts that integrate all three approaches. lVIotivation (Vol. 43, pp. 43-92). San Diego: Nevertheless, this seems like a potentially Academic Press. productive direction for theory develop Barsalou, L. W., Sloman, S. A, & Chaigneau, ment, and it will be interesting to see what S. E. (2005). The HIPE theory of func form theories of the conceptual system take tion. In Carlson, L. & van der Zee, E. (Eds.) in coming years. Representing fimctional features for language and space: Insights from perception, categori zation and development (pp. 131-47). Oxford: References Oxford University Press. Barsalou, L. vV., Simmons, W. K., Barbey, A. K., Anderson, J. R. (1976). Language, memory, and & Wilson, C. D. (2003). Grounding concep thought. Hillsdale, NJ: Lawrence Erlbaum tual knowledge in modality-specific systems. Associates. Trends in Cognitive Sciences, 7, 84-9i. Atran, S., Medin, D. L., & Ross, N. 0. (2005). The Barsalou, L. 'vV., Simmons, 'vV. K., Santos, A., cultural mind: Environmental decision mak Hamann, S. B., & Harenski, C. L. (2005). Word ing and cultural modeling within and across association and situated simulation in concep populations. Psychological Review, m, 7+1-76. tual processing. Manuscript in preparation. Barsalou, L. W. (1991). Deriving categories to Barsalou, L. W. & Wiemer-Hastings, K. (2005). achieve goals. In Bower, G. H. (Ed.) The psy Situating abstract concepts. In Pecher, D. & chology of learning and motivation: Advances Zwaan, R. (Eds.) Grounding cognition: The in research and theory (Vol. 27, pp. 1-64). San role of perception and action in memory, lan Diego, CA: Academic Press. guage, and thought (pp. 121j-63). New York: (1992). Frames, concepts, and conceptual fields. Cambridge University Pressc In Lehrer, A. & Kittay, E. F. (Eds.) Frames, Barsalou, L. W., Yeh, W., Luka, B. J., Olseth, K. L., fields, and contrasts: New essays in lexical and Mix, K. S., & Wu, L. (1993). Concepts and 254 BARSALOU
meaning. In Beals, K., Cooke, G., Kathman, transfer from entry-level to higher-level cate
D., McCullough, K. E., Kita, S., & Testen, D. gories. Connection Science, 121 143-62. (Eds.) Chicago Linguistics Society 29: Papers Cangelosi, A., Coventry, K., Rajapakse, R., Joyce, from the parasession on conceptual represen D., Bacon, A., Richards, L., & Newstead, S. :i tations (pp. 23-61). University of Chicago: (2005). Grounding language in perception: Chicago Linguistics Society. A connectionist model of spatial terms and Bartlett, F. C. (1932). Remembering: A study in vague quantifiers. In Cangelosi, A., Bugmann, experimental and social psychology. New York: .G., & Borisyuk, R. (Eds.) Modeling language, Cambridge University Press. cognition and action: Proceedings of the 9th Bechtel, W. & Abrahamsen, A. (2002). Neural Computation and Psychology Workshop Connectionism and the Mind: Parallel pro (pp. 47-56). Singapore: World Scientific. cessing1 dynamics1 and evolution in netzvorks. Cangelosi, A. & Riga, T. (2005). An epigenetic Cambridge, MA: Basil Blackwell. robotic model for sensorimotor grounding Bear, M. F., Connors, B. 'vV., & Paradiso, M.A. and grounding transfer. Under review. (2001). Neuroscience: Exploring the brain (2nd Carreiras, lvl., Garnham, A., Oakhill, J. V., & ed.). Baltimore: Williams & Wilkins. Cain, K. ( 1996). The use of stereotypical gen Biederman, I. (1981). On the semantics of a glance der information in constructing a mental at a scene. In Kubovy, M. & Pomerantz, J. R. model: Evidence from English and Spanish. (Eds.) Perceptual organization (pp. 213-53). Quarterly Joumal of Experimental Psychology, Hillsdale, NJ: Lawrence Erlbaum Associates. 49A, 63ir63. Bock, J. K. (1987). Co-ordinating words and Chambers, D. & Reisberg, D. (1992). What an syntax in speech plans. In Ellis, A. W. (Ed.) image depicts depends on what an image Progress in the psychology of language (Vol, 3, means. Cognitive Psychology, 24, 145-7+ pp. 337- :e. processing: Simulating semantic priming. comprehension. Psychological Review, 101, Cognitive Science, 23, 371-44. 371-<)5· :e, Damasio,A. R. (1989). Time-locked multiregional Haugeland, J. (1985). Artificial intelligence: The S. retroactivation: A systems-level proposal for very idea of it. Cambridge, MA: MIT Press. ·n: the neural substrates of recall and recognition. (1991). Representational genera. In Ramsey, id Cognition, 33, 25-62. W., Stitch, S.P., & Rumelhart, D. E. (Eds.) n, Donald, M. (1993). Precis of"Origins of the mod Philosophy and connectionist theory (pp. fo-89). ;e, ern mind: Three stages in the evolution of Hillsdale, NJ: Lawrence Erlbaum Associates. th culture and cognition." Behavioral and Brain Hayward, W. G. & Tarr, M. J. (1995). Spatial lan 1p Sciences, 16, 73cr .l Psychology, 3z, 97-127. . and the structure of concepts: A distributed :lge, Schwanenflugel, P. J. (Ed.) (1991). The psychology account of category-specific deficits. Brain lj of word meaning. Mahwah, NJ: Erlbaum. and Language, 75, 195-231. BARSALOU Wisniewski, E. J. (1997). When concepts combine. Author Notes Psychonomic Bulletin & Review, 4, 167-83. Yeh, W. & Barsalou, L. W. (2006). The situ This work was supported by National Science ated nature of concepts. American Journal of Foundation Grants SBR-9421326, SBR-9796200, Psychology, n9, 349-38+ and BCS-0212134 and by DARPA contracts BICA Zwaan, R. A. & Madden, C. J. (2005). Embodied FA8650-05-C-7256 and BICA FA8650-05-C-7255 sentence comprehension. In Pecher, D. & to Lawrence W. Barsalou. Address correspon Zwaan, R. (Eds.) Grounding cognition: The dence to Lawrence vV. Barsalou, Department of role of "perception and action in memory, lan Psychology, Emory University, Atlanta, GA 30322 guage, and thought (pp. 224-45). New York: ([email protected]) http://www. psychology. Cambridge University Press. emory. edu/cognition/barsalou/index.html).