The Prototype Theory of conceptual representation in large part owes its beginnings to Rosch and Mervis (1975), who, in the space of a couple of years, published a string of major papers laying out the empirical basis for the theory. The motivation for the theory came from a perceived crisis in philosophy and linguistics to do with defining the meaning of words. To the lay person, who has never worried too much about such things, the meaning of words is just given in the dictionary. The trouble is that most dictionary definitions are really only approximate or partial. The word "red," for example, is not defined by a fixed interval of the color spectrum but is the name for an imprecisely defined region with vague edges. The word "chair" could perhaps be defined as a movable object made for sitting on that stands on the floor and has a back. However, once again the actual use of the word tends in practice to allow for vagueness-designers continually create new objects for sitting on and new contexts in which to sit, so that it is often unclear whether they should be counted as chairs or not. The central insight of Prototype Theory is that word meanings and the conceptual classes that the words name are distinguished one from another not in terms of an explicit definition but in terms of similarity to a generic or best example. The concept red is the class of colors that are centered around a particular point on the spectrum that everyone tends to agree is the prototype red. Berlin and Kay (1969) reported that there was better agreement about the best examples of color terms than there was about the boundary between one color and another (e.g., between red and orange). The category of red things is therefore the category of things whose color is sufficiently similar to a prototypical red (and dissimilar from other prototypes). Similarly, there are concept representations for "chair" and "stool" and "bench" and "sofa," each of which is associated with a prototype example of the class. Objects are then classified on the basis of which prototype they are most similar to. Rosch, Simpson and Miller (1976) showed that people could readily learn novel categories based around prototypes (a point already demonstrated by Posner & Keele, 1968), and Rosch and Mervis (1975) analyzed a number of semantic categories, such as fruit, sport, or vehicle, to show that what members of the category had in common was not some set of defining features but a sufficient degree of resemblance to each other. In some of their writings, it is implied that the best example of the category, whatever that might be, would be the prototype. However, it quickly became clear that the prototype should better be considered as a more abstract, generic concept that was constituted from the different ways in which the category members resembled each other and differed from nonmembers. Unlike a best example, an abstract prototype allows for the representation of different possible values of relevant features such as that apples can be red, green, brown, or yellow or that furniture can be sat on, slept on, used for storing things, or provide a surface for supporting things. An apple that had all these colors or a piece of furniture that served all these functions would not necessarily be prototypical. Prototypes then are the centers of clusters of similar objects and prototype concepts form similarity-based categories. The center of the cluster is well established and agreed upon, but the boundary between one category and another may be subject to vagueness and disagreement. Talk of clusters with centers implies a spatial metaphor, and prototypes have often been discussed as points in similarity space. A mathematical exploration of the implications of this approach can be found in Gärdenfors (2000), and Osherson and Smith (1981) included a similarity space as part of their formalization of Prototype Theory. Spaces, however, have additional structural properties, which impose unnecessarily strict constraints on prototypes. Verbeemen, Storms, and Verguts (2004) have explored the degree to which natural categories can be represented in spaces (through multidimensional scaling) and concluded that at least for some semantic domains, a nonspatial similarity model provides a better fit. Following its introduction into cognitive psychology, Prototype Theory was also taken up enthusiastically by cognitive linguists, such as Ross (1973) and Lakoff (1987), and anthropologists such as Kempton (1978) and Randall (1976). Ross (1973), for example, proposed that the syntactic class NOUN in English is based around a prototype. He suggested a scale of "nouniness" associated with a hierarchy of syntactic acceptability in different contexts. The more nouny a word or phrase was, then the more contexts in which it would behave like a noun. A useful source of different views on the value of prototypes in linguistic theory can be found in Aarts, Denison, Keizer, and Popova (2004). While Rosch and Mervis provided overwhelming evidence for widespread prototype effects in semantic concepts and category learning, the development of the theory in psychology subsequently remained relatively underspecified. In one of the last chapters in the series, Eleanor Rosch (1978) discussed the theoretical underpinning of the data and warned that a distinction should be made between the empirical phenomena of prototype effects and any theoretical model that concepts are actually represented by prototypes. In fact, she doubted that the latter was the case. The purpose of this chapter will be to reexamine Prototype Theory and the evidence with which it is associated. One of the major difficulties with the theory may be that, with the early withdrawal of Rosch from the field, it has lacked a champion to develop and refine a working model of prototype representations, as new empirical results have been discovered. Thus, at various times, the theory has been criticized in many ways. For example, it is claimed that the theory lacks any way to represent the variability allowed on different dimensions within a category (e.g., the range of possible sizes of apples rather than just their average size). The theory is said not to be able to account for some categories having wider or more flexible boundaries than others [and, hence, is unable to explain why a sphere half-way in size between a basketball and a watermelon is more likely to be a watermelon than a basketball; (Rips, 1989)]. The theory is said to rely too heavily on statistical cue validity to determine feature weights (i.e., on the relative frequency of the feature for members and nonmembers of the category) and so to ignore causal dependencies among features such as that birds need their wings in order to fly. The theory is said to be circular in that no account is offered of why our attention is drawn to particular sets of features or particular sets of objects in the first place. In every case, the criticisms may be well-founded, but what has been lacking is a coordinated attempt to modernize the theory to incorporate mechanisms to deal with the failures. It is, of course, easy to find data that a model has no way of explaining, if the model was not created with those data in mind. However, one is then faced with a choice of discarding the model altogether or of adapting the model to fit the data. A notable exception to the lack of development of Prototype Theory has been the work on category learning of Don Homa and colleagues (Homa, 1984; Homa, Sterling, & Trepel, 1981) and Smith and Minda (2000). Both of these groups of researchers have generated valuable evidence that in classification learning paradigms, there are conditions under which abstraction of prototypes does occur. They have also developed precise quantitative models of how prototypes develop and are used in such learning situations. The question remains, however, whether the original aim of Prototype Theory-to provide an account of the natural concepts that we use to understand our everyday world and that serve to support the meanings of common nouns in natural language-can be met satisfactorily. The chapter therefore will focus on the original evidence on which Prototype Theory was based and will discuss which aspects of that evidence should be retained as central to the theory and which aspects may be less crucial. I will also use this opportunity to present new results relating to prototype effects and to reflect on some of the theoretical debates that surround the model. There is a nice irony here, in that the theory as applied to itself would suggest quite plausibly that "Prototype Theory" as a concept is itself a family of related concepts in which different importance might be attached to different assumptions of the theory. A prototype of Prototype Theory might be that presented by Rosch and Mervis, or that described in Hampton (1995), but other characterizations have been offered (Osherson & Smith, 1981). Leaving this irony aside, it is important first to try to capture the more essential characteristics of a prototype model, in order to consider how the central insights of the approach can be made consistent with evidence on the nature of conceptual representation. © 2006 Elsevier Inc. All rights reserved.
