Prototypes and Compositionality 89

antisceptical employment. By contrast, it’s their being complex that
primarily makes definitions interesting to psychologists and linguists.With
complex things, there’s always the hope that their behaviour can be
predicted from the behaviour of their parts; with primitive things, since
there are no parts, there is no such hope. In particular (for the linguists),
if words have definitions, then arguably words have the syntax of phrases
“at the semantic level”; so perhaps lexical grammar can be unified with
phrasal grammar. Likewise (for the psychologists), if lexical concepts are
tacitly structurally complex, perhaps they can be brought under the same
psychological generalizations that govern concepts that are manifestly
complex; if the concept BACHELOR is the concept UNMARRIED
MAN, then learning or thinking with the one can’t differ much from
learning or thinking with the other.2
So the definition theory was a fusion of disparate elements; in
particular, the idea that concepts are complex and the idea that their
constitutive inferences are typically necessary are in principle dissociable.
And, for better or worse, they have been coming unstuck in the recent
history of cognitive science. The currently standard view is that the
definition story was right about the complexity of typical lexical concepts,
but wrong to claim that complex concepts typically entail their
constituents. According to the new theory, it’s not the necessity of an
inference but its reliability that determines its relevance to concept
individuation.
How this is supposed to work, and why it doesn’t work the way that it’s
supposed to, and where its not working the way that it’s supposed to leaves
us in the theory of concepts, will be the substance of this chapter.
Statistical Theories of Concepts
The general character of the new theory of concepts is widely known
throughout the cognitive science community, so the exegesis that follows
will be minimal.
Imagine a hierarchy of concepts ordered by relations of dominance and
sisterhood, where these obey the intuitive axioms (e.g. dominance is
antireflexive, transitive and asymmetric; sisterhood is antireflexive, transitive,
and symmetric, etc.). Figure 5.1 is a sort of caricature.
Prototypes and Compositionality 89
2 The structural complexity of definitions was of some use to philosophers too: it
promised the (partial?) reduction of conceptual to logical truth. So, for example, the
conceptual truth that if John is a bachelor then John is unmarried, and the logical truth that
if John is unmarried and John is a man then John is unmarried, are supposed to be
indistinguishable at the ‘semantic level’.
The intended interpretation is that, on the one hand, if something is a
truck or a car, then it’s a vehicle; and, on the other hand, if something is
a vehicle, then it’s either a truck, or a car, or . . . etc. (Let’s, for the moment,
take for granted that these inferences are sound but put questions about
their modal status to one side.) As usual, expressions in caps (‘VEHICLE’
and the like) are the names of concepts, not their structural descriptions.
We continue to assume, as with the definition theory, that lexical concepts
are typically complex. In particular, a lexical concept is a tree consisting
of names of taxonomic properties together with their features (or
‘attributes’; for the latter terminology, see Collins and Quillian 1969),
which I’ve put in parentheses and lower case.3 In a hierarchy like 5.1, each
concept inherits the features of the concepts by which it is dominated.
90 Prototypes and Compositionality
. . . ARTEFACTS (+made objects)
. . . CHAIRS VEHICLES (+used for transport)
WHEELED VEHICLES
. . . CAR BICYCLE TRUCK
SPORTS CAR COUPE . . .
MACK ARTICULATED U-HAUL (+to rent)
(+self drive)
. . .
FIG. 5.1 An entirely hypothetical ‘semantic hierarchy’ showing the position
and features of some concepts for vehicles.
3 What, exactly, the distinction between semantic features and taxonomic classes is
supposed to come to is one of the great mysteries of cognitive science. There is much to be
said for the view that it doesn’t come to anything. I shall, in any case, not discuss this issue
here; I come to bury prototypes, not to exposit them.
(–flies) . . .
Thus, vehicles are artefacts that are mobile, intended to be used for
transport, . . . etc.; trucks are artefacts that are mobile, intended to be used
for transport of freight (rather than persons), . . . etc. U-Haul trucks are
artefacts that are mobile, intended to be rented to be used for transport of
freight (rather than persons), . . . and so forth.
The claims of present interest are that when conceptual hierarchies like
5.1 are mentally represented:
i. There will typically be a basic level of concepts (defined over the
dominance relations);
and
ii. There will typically be a stereotype structure (defined over the
sisterhood relations).
Roughly, and intuitively: the basic level concepts are the ones that receive
relatively few features from the concepts that immediately dominate them
but transmit relatively many features to the concepts that they immediately
dominate. So, for example, that it’s a car tells you a lot about a vehicle; but
that it’s a sports car doesn’t add a lot to what ‘it’s a car’ already told you.
So CAR and its sisters (but not VEHICLE or SPORTS CAR and their
sisters) constitute a basic level category. Correspondingly, the prototypical
sister at a given conceptual level is the one which has the most features in
common with the rest of its sisterhood (and/or the least in common with
non-sisters at its level). So, cars are the prototypical vehicles because they
have more in common with trucks, buses, and bicycles than any of the
latter do with any of the others.
Such claims should, of course, be relativized to an independently
motivated account of the individuation of semantic features (see n. 3).
Why, for example, isn’t the feature bundle for VEHICLE just the unit set
{+vehicle}? Well may you ask. But statistical theories of concepts are no
better prepared to be explicit about what semantic features are than
definitional theories used to be; in practice, it’s all just left to intuition.
That’s scandalous, to be sure; but fortunately it doesn’t matter a lot for
the issues that will concern us here. As we’re about to see, prototype
concepts and basic object concepts exhibit a cluster of reliably correlated
properties which allow us to pick them out pretty well even though,
lacking a theory of features, we have no respectable account of what their
basicness or their prototypicality consists in.
That concepts are organized into hierarchies isn’t, of course, anything
that definitional theories need deny.What primarily distinguishes the new
story about concepts from its classical predecessor is the nature of the glue
that’s supposed to hold a feature bundle together. Defining features were