Ideals Aren't Always Typical: Dissociating Goodness-of-Exemplar From Typicality Judgments

Ideals Aren’t Always Typical: Dissociating Goodness-of-Exemplar From Typicality Judgments Aniket Kittur (nkittur@ucla.edu) Keith J. Holyoak (holyoak@lifesci.ucla.edu) Department of Psychology, University of California, Los Angeles CA 90095 John E. Hummel (jehummel@psych.uiuc.edu) Department of Psychology, University of Illinois, Urbana Champagne, Urbana IL 61820 The standard way in which typicality is measured is through goodness-of-exemplar (GOE) judgments; for exam- ple, “How good an example is item A of category B?” This measure is so universally accepted that the concept of typi- cality is often introduced as synonymous with category goodness. For example, a classic paper on categorization asserts, “Instead of being equivalent, the members of a cate- gory vary in how good an example (or how typical) they are of their category” (Barsalou, 1985, p. 629). The reason that typicality and GOE are so often consid- ered equivalent is quite simple: in most studies of categori- zation they are indistinguishable. However (and, we argue, not coincidentally), most studies of categorization also use categories structured by central tendencies and represented by simple features. Abstract Items that are rated good examples of a category have gener- ally been assumed to be highly typical as well. However, most previous studies have used categories defined by simple features in which exemplar goodness and typicality are strongly related. We report a study using categories based on the relationships between features instead of the features themselves, allowing manipulation of relational ideals inde- pendent of featural central tendencies. Goodness-of-exemplar (GOE) judgments were based on relational ideals, whereas typicality judgments were based on a mix of ideals and featu- ral central tendencies. These results indicate that exemplar goodness and typicality can lead to two distinct forms of graded category structure, and should not be treated as equivalent. Keywords: categorization, category learning, similarity, rela- tions, typicality, goodness-of-exemplar Central Tendencies versus Ideals Introduction Central tendency views of category structure have been dominant for the last three decades. The exemplars of a category structured by central tendency are considered bet- ter members the more similar they are to the “center” of the category. This similarity measure may be discrete, such as the number of shared and unshared properties between ex- emplars represented as lists of features (Rosch & Mervis, 1975), or continuous, such as the distance from a dot pattern prototype (Posner & Keele, 1970). It may be calculated by the distance between two items in a hierarchy (Lynch, Coley, & Medin, 2000), by the distance between two points in a stimulus space (Ashby & Gott, 1988), or by many other methods. However, a fundamental assumption of a central tendency structure is that the further from the “center” an exemplar gets, the worse an example of the category it be- comes. Importantly, we will refer to “central tendencies” here as those metrics that depend only on the distribution of individual features. These may include such metrics as fa- miliarity, frequency of instantiation, shared features, and distance from a prototype or exemplars 1 . All of these can be calculated without reference to other categories, goals, or relations. However, central tendencies do not appear to define all categories. Armstrong, Gleitman, and Gleitman (1983) One of the most robust findings in categorization research is the graded structure of categories. Every member of a category is not considered an equally good example of the category; instead, items lie on a spectrum of category good- ness (Rips, Shoben, & Smith, 1973; Rosch & Mervis, 1975). For example, a robin is considered by American under- graduates to be a better example of a bird than is an ostrich or a penguin. This graded goodness-of-example effect is known as “typicality”, and there is a large body of work supporting its influence on categorization. Object classification speed increases with typicality (Rips et al., 1973); for example, people are fast to affirm that a robin is a bird, but slower to affirm that a chicken (a less typical bird) is. More typical items are generated before less typical ones (Mervis, Catlin & Rosch, 1976). People are more likely to extend infer- ences when the source of the inference is a typical category member rather than an atypical one (Rips, 1975). Category learning is faster when typical rather than atypical items are presented earlier in the sequence (Mervis & Pani, 1980; see also Posner & Keele, 1968). In fact, the prevalence of typi- cality effects in categorization has led some researchers to say that “if one compares different category members and does not find an effect of typicality, it suggests that there is something wrong with – or at least unusual about – the ex- periment” (Murphy, 2002, p. 24). We do not try to distinguish between prototype and exemplar theories here, as both predict that items further from the central tendency of the category will be considered worse exemplars.