Challenging assumptions about concepts
Post on 23-Nov-2016
Cognitive Development, 8, 169-180 (1993)
Challenging Assumptions About Concepts
Lawrence W. Barsalou University of Chicago
A science progresses by challenging its assumptions. As assumptions come and go, they represent current understandings, simplify analysis, and sometimes exist simply because believing otherwise would be inconceivable or present intractable problems. Within existing assumptions, theories develop and experi- ments are performed. If these accomplishments satiate the needs for knowledge, there is no reason to press further. Yet, if outstanding issues remain, as they usually do, current assumptions may block their resolution, and the search for new assumptions often begins. The study of concepts across the cognitive sci- ences, from philosophy and linguistics to psychology and computer science, has rested on some fairly well-accepted assumptions for some time. Clearly, much progress has been made in this context, as the explosion of successful research on concepts over the last 20 years illustrates. Nevertheless, increasing numbers of researchers are challenging these assumptions, as do Jones and Smith (1993) in their provocative, insightful, and compelling lead article in this issue of Cogni- tive Development. In my commentary, I similarly question these assumptions, posing four questions that those of us who study concepts might ask ourselves from time to time: (a) Do coherent conceptual cores exist in long-term memory? (b) Do abstract propositions constitute conceptual cores? (c) Do concepts in long- term memory control behavior? (d) Is the primary purpose of developing and using concepts to taxonomize the environment?
The answers to these questions may in fact be that there are coherent concep- tual cores in long-term memory, built from abstract propositions, that control behavior, as people taxonomize the environment. Yet for such answers to be convincing, researchers must transform these assumptions into hypotheses and marshall explicit evidence for them, rather than, as at present, embracing them implicitly, simply because of their familiarity, intuitive appeal, and convenience. As we shall see, these assumptions face strong challenges, both empirically and theoretically, and there is good reason to believe that an alternative state of affairs exists.
Work on this article was supported by funding from the Army Research Institute, MDA 903-90- K-0112.
Correspondence and requests for reprints should be sent to Lawrence W. Barsalou, Department of Psychology, 5848 S. University Ave., University of Chicago, Chicago, IL 60637.
Manuscript received September 30, 1992; revision accepted January 13, 1993 169
170 Lawrence W. Barsalou
ARE THERE COHERENT CONCEPTUAL CORES IN LONG-TERM MEMORY?
To place this first issue in perspective, consider the two extreme forms that the representation of a concept in long-term memory could take: At the incoherent extreme lie exemplar representations of categories. At the coherent extreme lie consistent and complete logical systems of categories. Exemplar representations of categories are potentially incoherent, because such systems, in their standard form, make no attempt to integrate exemplars or to develop generalizations across them. Instead, the representation of a concept is fragmented into indepen- dent exemplar memories acquired during different processing episodes with the category, with no generalizations across exemplars providing integration. As a result, these representations have no conceptual cores. Inconsistencies can occur, if different exemplars contain contradictory information about the category. In- completeness can occur, if existing exemplars do not provide an error-free algo- rithm for classifying all new exemplars.
In contrast, consistent and complete logical systems of categories do not exhibit inconsistencies, because deductive rules ensure that no piece of informa- tion in such a system contradicts any other piece of information. Nor do these systems exhibit incompleteness, because they provide error-free algorithms for classifying all new exemplars (e.g., necessary and sufficient conditions for cate- gory membership). Maintaining a consistent and complete system of category knowledge is effortfui, complicated, and sometimes intractable, as work on truth maintenance systems and the frame problem in artificial intelligence illustrates (Brown, 1987; McCarthy & Hayes, 1969).
Between these two extremes lie many other possibilities, and probably the true state of affairs. One possibility is the core-plus-identification view, which can be construed as a concatenation of the two extremes just reviewed (e.g., E.E. Smith & Medin, 1981). On the one hand, cores contain consistent and complete criteria for category membership that are effortful to apply. On the other hand, identification procedures contain heuristics for recognizing category members that are implemented easily, but that may be inconsistent and incomplete. Alter- natively, there may be no sharp distinction between core and peripheral proper- ties, with properties simply varying continuously from highly diagnostic or important to weakly diagnostic or unimportant (e.g., Hampton, 1979, 1987; McClelland & Rumelhart, 1985). A weak form of coherence might arise in these theories when central properties are intercorrelated with one another across ex- emplars, thereby becoming active across contexts as a set, as more peripheral properties come and go.
Thus far, there appears to be no compelling evidence for or against any of these views. Each position has its adherents, yet direct evidence is sparse. A number of critical issues must be addressed empirically and theoretically before we can answer this question more definitively. First, we must determine whether
Challenging Assumptions About Concepts 171
natural concepts really have conceptual cores, or whether the appearance of cores arises from other sources of knowledge. As Jones and Smith suggest, people may simply have a metatheory about categories in general, which contains the belief that any category should have a core. In the induction task, people may use this metatheory about categories to infer that the category under consideration has a core, even though they have no knowledge about its content. This inference could in turn cause a category name to control induction, because people believe that two entities having the same name also share the same core and should therefore be sorted together. If this metatheoretical belief about the general existence of cores controls induction, it does not follow that people know con- ceptual cores for specific categories.
Alternatively, a metatheory about linguistic expertise could control induction (Putnam, 1973, 1975). Perhaps subjects in induction experiments perceive the experimenter as a linguistic expert and believe that the experimenter's use of category names reflects his or her expertise. When the experimenter refers to a perceptually dissimilar exemplar by the name of the critical category, subjects may infer from their metatheory about linguistic expertise that this exemplar is the one most likely to possess the induced property.
Still another possibility is that people know conceptual cores only for very general categories, such as ontological kinds (Keil, 1979). For example, people might know cores for animal, artifact, location, and so forth. In induction tasks involving more specific concepts, such as skunk and chair, people may infer properties from their more general superordinates via inheritance. Because these core properties reside outside of the specific concepts under consideration, claiming that these concepts have cores is inaccurate. Even though concepts may inherit core information, only a few concepts actually have explicitly represented cores, and most do not.
As these examples illustrate, people may not know conceptual cores for specific categories, yet they may act as if they do because of other knowledge they possess. Before we conclude that specific concepts have cores, we need direct evidence for such cores and evidence against the possibility that they arise from other sources of knowledge.
If we discover that conceptual cores exist in specific concepts, a second set of issues arises concerning their acquisition and maintenance. How do people acquire cores? Discovering definitional properties for a concept can require a tremendous amount of computational effort, perhaps more than people are capable of performing (Barsalou & Hale, 1992). Furthermore, how do people maintain cores? Avoiding inconsistency in the face of a changing world raises tremendous computational problems; and it would perhaps be surprising, given everything we know about the limitations on human rationality, to find that people can avoid inconsistency all that well. Maintaining consistency could require examining everything one knows about a category on encountering each new exemplar, and it would be remarkable if people were this thorough in
172 Lawrence W. Barsalou
maintaining their knowledge. As further cause for skepticism, we have observed specific individuals in our laboratory producing contradictory information when defining the same concept on two different occasions (Barsalou, Sewell, & Spindler, 1992).
It may be that people try to establish some generality and consistency in their concepts of categories, but only succeed partially. For example, people may establish coherent generalizations across subsets of exemplars but not across all exemplars (Medin & Ross, 1989), or they may establish coherent generalizations that apply only in specific contexts. If one assumes that just a small subset of knowledge is ever activated on a given occasion for a category, then coherence may only be established within these activated subsets, or between several sub- sets processed sequentially. As a result, there may be many "patches" of local coherency within the greater knowledge of a category, with no overarching generalizations establishing consistency across all of it. Perhaps the only place that such broad generalizations occur successfully are in formal approaches to knowledge, as in science, although inconsistencies often occur there as well.
Finally, a third set of issues concerns the use of cores. Are cores used every time a concept is processed, or just sometimes? Perhaps cores are used only when the costs for making a categorization error are higher than the costs of applying the core. When the costs of a categorization error are less than the costs of applying the core, people may often use more heuristic methods that are easier to apply. For example, while wandering across a university campus, one categor- izes students and faculty based on appearance prototypes, simply because the cost of checking university identification cards is higher than the cost of making an error.
If most defining information associated with a category is only used occasion- ally, why should we distinguish it with the term "core'?" Why are the expert and least defeasible criteria for determining membership thought to constitute the most important information that the average person knows about a concept? Why isn't the concept's function more central (Nelson, 1974)'? Because functions are probably of significantly greater importance than definitions in daily life, why don't they constitute conceptual cores (e.g., the function of water versus its molecular structure; cf. Malt, 1992)'?
DO ABSTRACT PROPOSITIONS CONSTITUTE CONCEPTUAL CORES?
As Jones and Smith note, theorists often assume that abstract propositions, rather than perceptual memories, constitute the cores of concepts. The primary founda- tion of the propositional view in cognitive science stems from the discovery that people forget surface structure in language and pictures quickly, retaining only their gist (e.g., Mandler & Ritchey, 1977; Sachs, 1974). To represent gist, theorists turned to propositional representations, because multiple surface repre-
Challenging Assumptions Abou! Concepts 173
sentations (e.g., the active and passive forms of a sentence) can reflect the same underlying propositions (e.g., Anderson & Bower, 1973; Kintsch, 1974). Anoth- er reason for adopting propositions is to represent abstract concepts. Because propositions are abstract amodal representations that bear no perceptual resem- blance to their referents, their predicate-calculus-like structures can represent the content of any concept, not just perceptual ones.
Nevertheless, the propositional view of concepts faces serious problems (Bar- salou, 1993). The first is that no one has ever provided direct evidence that anything like predicate-calculus expressions lie at the heart of cognitive represen- tations. Instead, we have only adopted these representations because they pro- vide a convenient formal language for expressing what theorists believe about the content of concepts. A second problem is that we have no accounts of how propositional representations arise in the cognitive system, either innately or through experience. We haven't the faintest idea of how biological mechanisms could produce abstract propositions, nor of how information processing mecha- nisms might transduce propositions from perceptual experience.
A related difficulty is symbol grounding, as illustrated by Searle's (1980) Chinese room problem (also see Harnad, 1987, 1990; Huttenlocher, 1976; Hut- tenlocher & Higgins, 1978). Because propositional representations bear no per- ceptual similarity to their referents, the relations between these representations and their referents are arbitrary. If no account is given of how propositional representations are related to their referents, the result is a system that has no meaning or true understanding of its world.
Because we have no direct evidence for propositional representations, because we can not tell a coherent story about their origins in the cognitive system, and because we have no account of how propositions relate to the environment, we should be at least somewhat skeptical of them, and we might even consider looking for alternatives. Elsewhere, I have outlined an alternative view in which all conceptual representations arise from perception (Barsalou, 1993). This theo- ry makes several departures from what might typically be viewed as the percep- tual approach to knowledge. First, the perceptual representations that underlie knowledge are not holistic exemplars or analog "snapshots" of perception. In- stead, they are small components of perceptions extracted analytically from larger experiences via selective attention (e.g., the shape, texture, or color of an object). Second, such components are extracted not only from perceptions of the environment, but also from introspection of subjectively-experienced mental events. As a resialt, perceptual components develop for cognitive operations (e.g., retrieval, c...