Discussion

On prototypes as defaults (Comment on Connolly, Fodor, Gleitman and Gleitman, 2007)☆,☆☆

In an interesting contribution to research on conceptual combination, Connolly, Fodor, Gleitman, and Gleitman (2007) (CFGG) tested a hypothesis that they identified as a common assumption of prototype models of conceptual combination, ‘the crucial default to the stereotype prediction’ (DS). Defaulting to the stereotype consists in assuming, barring information to the contrary, that the prototype corresponding to an adjective noun combination (AN), fully inherits the properties of the prototype corresponding to the noun (N) of that combination. For instance, the prototype corresponding to uncomfortable sofa should share all the properties of the sofa prototype, except for those properties that have to do with comfort. In contradiction of this hypothesis, they reported an experiment in which people’s willingness to accept that a property is true of the members of some class tended to be lower when the concept was modified (uncomfortable sofas have backrests) than when it was unmodified (sofas have backrests). According to CFGG, since all prototype models of conceptual combination require DS to be true, the empirical demonstration that DS fails should be taken as important evidence against the correctness of such models.

Prior to reporting their experimental results, CFGG offered an argument to the effect that DS is in any case non-optimal and ‘invites indefinitely many bad bets’. Thus, according to CFGG, prototype theory embraces a misleading strategy, liable to lead people to make false inferences about the properties of complex concepts. Together, the empirical demonstration that people do not follow DS and the argument that DS is a bad strategy provide sufficient reason to favor some form of the ‘classical’ model over prototype models as the best approach to conceptual combination. The model they favor involves the assumption that (in the default case) modifiers and nouns correspond to concepts which are combined without being transformed into a structure having both of the original concepts as parts, so that the concept LITHUANIAN HAIR corresponds simply to a structure having the concepts LITHUANIAN and HAIR as parts. The classical combinatorial process involves no propagation of the prototypical/stereotypical properties of HAIR to LITHUANIAN HAIR although world knowledge may be evoked post-combination in order to make informed judgments about the likely properties of the complex concept.

Our paper aims to shed additional light on this topic, first by drawing out the implications of some recently published data that has explored and extended CFGG’s empirical result, and then by considering a number of important points in their paper. We disagree with them on several counts. First, we show that the results they report greatly exaggerate the extent to which subjects actually refrain from using the DS strategy. In fact when given the opportunity, most of the time people choose to say that a property is equally likely of a modified as of an unmodified noun, even when two atypical modifiers are used. Second, many of the prototype models that they claim to discredit do not embody the DS strategy at all, certainly not in the sense of rigidly defaulting to the stereotype regardless of the influence of knowledge-based inferences. Third, the results they report, contrary to undermining a central property of prototype approaches to conceptual combination, actually support a thesis that all these approaches do have in common – that the prototypes corresponding to complex expressions depend on the prototypes corresponding to their parts. Fourth, their argument that DS is an inefficient strategy is weak, and an alternative principled argument is offered for why in fact one should use this strategy. Our argument goes further in justifying why confidence in the inheritance of default attributes should be moderated by the typicality of the modifier – a point in their data that CFGG do not directly address.

We will argue for each of these points separately in the following sections. The final section argues that their results provide little reason to prefer the classical approach over the prototype approach to conceptual combination.