Graded similarity in free categorization

Abstract

Similarity has long been regarded as a major determinant of human categorization. Surprisingly, much research has shown that when people are asked to construct their own categories they rarely do so on the basis of overall similarity, instead categorizing on the basis of a single feature or dimension of the objects. This article reports five experiments that manipulate the proportion of parts shared by two structurally alignable objects to determine whether similarity would have a graded effect on free categorization. Increasing the proportion of shared features increased both the rated similarity of a given pair of objects and the probability of assigning them to the same category. Interestingly, the shape of the two similarity functions differed, with rated similarity increasing linearly with the proportion of shared features while the probability of assigning the objects to the same category increased superlinearly (exponentially). This difference is discussed in terms of Shepard's (1987) model of generalization, which predicts that any monotonic increase in perceived similarity will result in an exponential increase in the probability of generalization. Overall, these results provide a strong demonstration of similarity-based free categorization, and the particular form of that relationship provides useful information regarding the underlying cognitive processes involved.

Keywords: Alignment; Categorization; Category construction; Concepts; Generalization; Unsupervised learning.