Knowledge structures and linear separability: integrating information in object and social categorization

Abstract

These experiments were designed to determine if the naturalness of abstract category structures varies with content domain. Specifically, the degree to which linear separability constrains categorization was investigated in object and social domains. Linearly separable (LS) categories are categories that can be perfectly partitioned on the basis of a weighted, additive combination of component information. Across a wide variety of stimulus materials and classification tasks LS structures were found to be more compatible with social than object materials. In sorting tasks, participants were more likely to sum characteristic features and form LS categories with social materials. In learning tasks, LS structures were easier to learn with social materials but nonlinearly separable structures were easier to learn with object materials. This interaction between category structure and content domain was attributed to differences in the types of knowledge and integration strategies that were activated. In object conditions, strategies that were inconsistent with adding independent features were observed (e.g., focusing on single dimensions, using configural properties, and relying on analogy). In social conditions, however, summing the evidence and learning LS structures appeared to be a natural strategy. It was concluded that the structure of knowledge varies with domain, and consequently it will be difficult to formulate domain general constraints in terms of abstract structural properties such as linear separability. Differences between object and social categorization systems are discussed.