Abstract

Kinds of Kinds: Sources of Category Coherence Kenneth Jeffrey Kurtz ([email protected]) Dedre Gentner ([email protected]) Department of Psychology, Northwestern University, 2029 Sheridan Rd Evanston, IL 60208-2710 USA Abstract A fundamental question in the study of concepts is what makes sets of examples cohere as categories. We present results of three studies designed to compare standard taxonomic categories with categories that take their meaning from relationships extending outside of the individual example. An exemplar generation task is used to differentiate relational categories from taxonomic kinds and to compare possible subtypes of extrinsically cohering categories based on goals or thematic contexts. Results provide strong support for the intrinsic— extrinsic distinction and reveal signatures of underlying organization among the types of categories investigated. Introduction Categories play a fundamental role in cognition. The internal structure of categories supports numerous functions including classification, prediction, and reasoning. Categories give rise to an extension: the set of examples in the world that are members. The coherence of a category is the meaningful basis according to which these members go together. One traditional view holds that the correlational structure of the environment determines category coherence due to systematic patterns of within-category similarity and between-category difference (Rosch & Mervis, 1975). Murphy and Medin (1985) propose the theory-based view that challenges the idea that similarity itself explains category coherence. ‘Respects’ for similarity (i.e., a basis for the selection of features and weights) must be specified in order for concepts to exist as groups of like examples. Additionally, they argue that category representations are richer than lists of features and must include relationships that hold within and between examples of categories. The tension inherent in the need for a constrained, yet rich basis of category coherence poses a continued challenge to theorists (Goldstone, 1994). A useful source of inspiration is structural alignment theory—which has proven successful as an account of comparison processes such as similarity and analogy (Gentner, 1983; Markman & Gentner, 1993; Gentner, Rattermann, & Forbus, 1993). This framework offers a perspective for addressing the question of category coherence in a manner in keeping with the theory view. Respects for similarity can arise from the process of aligning corresponding predicates of two structured representations (Gentner & Markman, 1997; Medin, Goldstone, & Gentner, 1993; Markman & Wisniewski, 1997). Relational similarity drives the alignment process and largely determines the quality of a match. In the same way that structural alignment theory looks to shared relations (more than attributes or objects) to explain similarity, we can look to relationships between objects as a source of category coherence. While theory-based categories may cohere around intrinsic relationships (like the causal link between genetic material and physical features), we focus here on relationships extending beyond the individual example. For instance, the category barriers consists of examples that conform to the relationship: BLOCKS (X, Y). Barr and Caplan (1987) distinguish between the intrinsic properties of a category which are true of an example in isolation versus extrinsic features which hold only in relation to other objects. As an alternative to category members bound together by common intrinsic structure (relations or attributes), category coherence can be derived from relations extrinsic to individual examples. The extreme case of extrinsic coherence is relational categories like barriers— members cohere based on fulfilling a core relationship. The roles of X and Y in the blocking relation can be filled by anything—so as long as the relationship holds, membership is secure. The examples of a relational category may have few or no intrinsic properties in common with one another. In this sense, relational categories are akin to analogies. Both ‘prison bars’ and ‘raging river’ are members of the category barrier, despite their sharing no intrinsic similarity. Another case in which the category coherence is extrinsic is Barsalou’s (1983, 1985) ad-hoc or goal- derived categories which are organized around ideals (properties that optimally promote goal resolution) rather than central tendency. Again, categories such as things to take out of the house in case of fire violate the correlational structure of the environment since member examples have few properties in common. Goldstone (1995) makes a useful distinction between default and directed similarity; where the former is the basis of graded structure and broad inferential power of taxonomic categories, while the latter is the focal, context-specific sense of similarity underlying ad-hoc categories or analogical relationships. Categories may also be grounded by properties beyond the individual example that are not specific relationships. As an example, consider items associated