Similarity in context

variety of cognitive acts, ranging from problem solving to categorization to memory retrieval. If William James (1890/1950) was correct in stating that “this sense of Sameness is the very keel and backbone of our thinking” (p. 459), we might expect similarity judgments to be stable, reliable, and less flexible than the cognitive acts that depend on them. There is considerable research, however, that undermines this assumption (e.g., Medin, Goldstone, & Gentner, 1993). The three experiments to be reported here provide evidence that assessing the similarity of two things depends on the context of the judgment in several important and specific ways. The contextualized nature of similarity is shown by demonstrating systematic violations of several assumptions associated with standard models of similarity. Furthermore, the results show that “context” is not limited to the items actually present in a choice situation; two compared objects may create, or “recruit,” their own context, which in turn influences judged similarity. Contextual Effects in Similarity The notion of context dependence stands in contrast to what we will call fixed-set approaches to similarity. In this view, similarity is computed by integrating evidence from a fixed set of features or dimensions. Each of the things to be compared is first assigned a representation, in terms of features (Tversky, 1977), or values along continuous dimensions (Carroll & Wish, 1974; Shepard, 1962). The features or dimensions may also be assigned weights that indicate their salience. The representations are then compared in terms of overlap (Tversky, 1977), distance in psychological space (Caroll & Wish, 1974), or transformational distance (Imai, 1977; Wiener-Ehrlich & Bart, 1980). Importantly, the featural or dimensional representations are determined before the comparison process takes place, although in some methods, most notably Tversky’s contrast model, the context of a comparison may influence the weights assigned to the features of an object’s representation. The assumption that fixed-object representations are the inputs to a similarity computation is useful in several ways. It is a necessary precondition for many of the similarity techniques to operate. For example, in multidimensional scaling (MDS), objects are represented by points in multidimensional space. Context or intrinsic variability may alter an object’s position in the space (Ennis, 1992), and attentional changes may stretch or shrink the space in one or more dimensions (Nosofsky, 1986). However, if one allows that different object properties are systematically considered when the object is compared with different objects, then consistent and appropriate point locations for