110 Prototypes and Compositionality

to
Prototypes and Compositionality 109
sensory concepts. It’s sometimes suggested that they illustrate the presence
of an “iconic” element in concepts like RED (see the discussion above of
Jackendoff 1992). Maybe ‘red’ means something like ‘similar in respect of
colour to this’ where the ‘this’ ostensively introduces a red sample. The
trouble with taking this line, however, is that the pattern RED and the like
exemplify actually appears to be quite general: lots of lexical concepts for
which definitions are very hard to find nevertheless appear to enter into the
same sort of “one way” entailments that hold between ‘red’ and ‘colour’.
It’s plausible that ‘dog’ means animal, but there doesn’t seem to be any F
(except DOG) such that ‘F + ANIMAL’ means dog. ‘Chair’ means
furniture, but what and FURNITURE means chair? Notice that it won’t
do to appeal to ‘iconic elements’ in these non-sensory cases. Maybe ‘red’
means ‘similar in colour to this’, but ‘dog’ doesn’t mean ‘similar in X to
this’ for any X that I can think of except doghood. It appears that, contrary
to traditional Empiricist doctrine, many lexical items are not independent
but not definable either; ‘red’ entails ‘colour’ but can’t be defined in terms
of it.
A natural way to accommodate the residuum problem is to allow that
some content-constitutive inferences don’t arise from definitions after all.
It’s not that RED entails COLOUR because the definition of ‘red’ is
COLOUR & F; rather, RED just entails COLOUR full stop. Following
the historical usage, I’ll call a principle of inference that institutes a ‘one
way’ relation of entailment between lexical concepts a “meaning
postulate”. Rules of lexically governed inference that happen to be
biconditional, like ‘bachelor « unmarried man’, have no special status
according to the theory that meaning postulates are what license lexically
governed inferences. This version of Inferential Role Semantics is therefore
weaker than the definitional account; the latter allows a lexical concept to
enter into constitutive inferential relations only if it is definable.
From our perspective, the important consequence of this liberalization
is that it disconnects the question whether an inference from C to C1 is
content-constitutive from the question whether C1 is a syntactic part of C.
Notice that it was only because definitions were required to be biconditional
that they could be viewed as exhibiting the structural description
of a concept. UNMARRIED MAN can’t be the structural description of
BACHELOR unless ‘BACHELOR’ and ‘UNMARRIED MAN’ denote
the same concept. But BACHELOR and UNMARRIED MAN can’t be
the same concept unless ‘BACHELOR «UNMARRIED MAN’ is true.
Detaching the question whether RED entails COLOUR from the
question whether COLOUR is a constituent of RED has its virtues, to be
sure.We’ve been seeing how weakly the empirical evidence supports claims
for the internal structure of lexical concepts. Meaning postulates allow
110 Prototypes and Compositionality
one to give up such claims while holding onto both ‘‘red’ means colour is
analytic’ and ‘you don’t have RED unless you know that red is a colour’.
On the meaning postulate story, RED ® COLOUR could be meaningconstitutive
even if neither RED nor COLOUR have any internal
structure; i.e. even if it’s atomic.
But no free lunch, of course.We started out this chapter by remarking
that one of the nicest things about the definition story was that it explains
an otherwise striking and perplexing symmetry between the metaphysics
of meaning and the metaphysics of concept possession: the very inferences
that are supposed to define a concept are also the ones you have to accept in
order to possess the concept. This really is striking and perplexing and not
at all truistic; remember, it isn’t (can’t be) true of all necessary inferences—
or even of all a priori inferences—that they determine the conditions for
possessing the concepts involved in them. Well, the theory that concepts
are definitions gets this symmetry for free; it follows from the fact that
definitions relate concepts to their constituents. If C is literally a part of C1,
then of course you can’t have C1 unless you also have C. Notice that this
explanation turns on precisely the idea that meaning postulates propose to
abandon: viz. that the content-constitutive inferences are the ones that
relate a concept to its parts.
In short, if you are independently convinced both that there are
meaning-constitutive inferences and that most lexical concepts behave like
primitives, you’ve got a residuum problem to which meaning postulates
may indeed offer a solution. But at a price, since the solution weakens the
architecture of your overall theory: it breaks the connection between the
structure of a concept and its possession conditions.
Partee has tried bravely to make a virtue of this necessity:
Meaning postulates might be a helpful tool . . . since they make the form [sic] of
some kinds of lexical information no different in kind from the form of some
kinds of general knowledge. That would make it possible to hypothesize that the
very same ‘fact’—for example, whales are mammals—could be stored in either of
two ‘places,’ a storehouse of lexical knowledge or a storehouse of empirical
knowledge; whether it’s part of the meaning of ‘whale’ or not need not be fixed
once and for all. (1995: 328)
But it is inadvisable for a theory to recognize degrees of freedom that
it is unable to interpret. Exactly because meaning postulates break the
‘formal’ relation between belonging to the structure of a concept and being
among its constitutive inferences, it’s unclear why it matters which box a
given such ‘fact’ goes into; i.e. whether a given inference is treated as
meaning-constitutive. Imagine two minds that differ in that ‘whale ®
mammal’ is a meaning postulate for one but is ‘general knowledge’ for the
Prototypes and Compositionality 111
other. Are any further differences between these minds entailed? If so,
which ones? Is this wheel attached to anything at all?
It’s a point Quine made against Carnap that the answer to ‘When is an
inference analytic?’ can’t be just ‘Whenever I feel like saying that it is’.
Definition versions of IR Semantics can hold that an inference is analytic
when and only when it follows from the structure of a concept. If the
meaning postulate version has an alternative proposal on offer, it’s not
one that I’ve heard of.