Matching performance and the similarity structure of the stimulus set

To compare two general approaches toward understanding how objects are classified, this article explores performance in a series of studies that required matching visually presented letters. One general approach has stressed the analysis of details or elements of the input pattern. The other approach has focused on the relationships between the various elements and objects that the observer might expect to occur in the task. In empirical support of the first approach, many researchers have observed that the time it takes a subject to report that two visually presented letters have the same name is less when these letters are physically identical (e.g., A-A) than when they are physically different (e.g., A-a). This relative ease of matching physically identical letters has been attributed to a visual process that matches the letters on the basis of their physical characteristics. In support of the second approach, the studies in this article replicate this finding, but only for data averaged across letter parts. The data from individual letter pairs do not reflect this temporal hierarchy. Individual letter performance for an entire set of letters has not previously been reported, and the importance of analyzing fine structure in data is stressed. It is shown that physical identity matches can be reliably faster or slower than name identity matches. The similarity structure of the total stimulus set reliably predicts which result will occur. Within the limits of the variables studied, this conclusion is shown to be independent of the criterion for the type of match the subject is asked to make (physical identity or name identity) and of the temporal and physical separation of the letters. The readiness with which the presented stimuli can be discriminated from other members of the stimulus set controls performance. For example, as Experiment 1 shows, it is easy or difficult to report o-o as "same" depending on what other letters are used in the task. Lockhead's holistic-discriminability model provides a framework for interpreting these demonstrations that matching performance does not depend only on the stimuli physically present. Performance depends on the similarity or discriminability between the presented stimuli and other stimuli the observers know might be presented.