Thus we see again the characteristic non-Boolean interaction of conditions to for ma cluster concept.
In Jackendoff (1983) I called a syste mof conditions of this sort a“preference rule system”; Lakoff (1987) calls it (one
aspect of) an“idealized cognitive model”; theframesof Minsky (1975) have similar effects. Concepts organized this
way, like Tarskian concepts, are combinations of conditions. They differ from Tarskian concepts in that the conditions
are combined differently, namely by“ψ-or.”Thus a proper theory of word meanings must go beyond traditional
philosophical assumptions. (Default logic comes closer to capturing the effect of such conditions, which function as
default values where there is no evidence to the contrary, for instance in (10a).)
Appealing to the psychological goals of conceptualist semantics, such enrichment of the theory proves to be plausible
on broader grounds. The manner in which conditions interact in cluster concepts is central in visual perception
(Wertheimer 1923; Marr 1982), in phonetic perception (Liberman and Studdert-Kennedy 1977), in musical cognition
(LerdahlandJackendoff1983),andinGriceanimplicature(BachandHarnish1979)(seeLerdahland Jackendoff1983:
ch. II for discussion of all of these). Moreover, such an interactioncan be plausibly instantiated in the brain, where the
firing of a neuron is normally not a rigid function of other neuralfirings (like a logical conjunction), but rather a
composite of many excitatory and inhibitory synapses of varying strengths. Thus cluster concepts, even though
unusual in a logical framework, are quite natural in a psychological framework.
A different interpretation of cluster concepts appears to be more widespread in psychology. Laurence and Margolis
(1999: 27), for instance, say“According to the Prototype Theory [e.g. Smith et al. 1988], most concepts—including
most lexical concepts—are complex representations whose structure encodes a statistical analysis of the properties
their members tend to have.... [Application is a matter of satisfying a sufficient number of features, where some may
be weighted more significantly than others.” The present analysis is not inherently statistical, although learning
doubtless involves some statistical correlation of attributes. However, the differences between the two approaches
remain to be explored.