(forthcoming in The Routledge Encylopedia of Philosophy)
According to Conceptual Role Semantics ("CRS"), the meaning of a representation is the role of that representation in the cognitive life of the agent, e.g. in perception, thought and decision-making. It is an extension of the well known "use" theory of meaning, according to which the meaning of a word is its use in communication and more generally, in social interaction. CRS supplements external use by including the role of a symbol inside a computer or a brain. The uses appealed to are not just actual, but also counterfactual: not only what effects a thought does have, but what effects it would have had if stimuli or other states had differed. The view has arisen separately in philosophy (where it is sometimes called "inferential," or "functional" role semantics) and in cognitive science (where it is sometimes called "procedural semantics"). The source of the view is Wittgenstein (1953) and Sellars, but the source in contemporary philosophy is a series of papers by Harman (see his 1987) and Field (1977). Other proponents in philosophy have included Block, Horwich, Loar, McGinn and Peacocke (1992). In cognitive science, they include Woods (1981) and Miller and Johnson-Laird (1976). (See references in Block, 1987.)
There are two quite different projects that go by the name 'semantics'. One, which we might call linguistic semantics, deals with the meanings of particular expressions in particular languages and how they fit together to make up meanings of larger expressions, for example, how the meaning of `John hopes angels exist' is related to the meaning of `angels exist'. The second project, metaphysical semantics, is one of investigating the fundamental nature of meaning, especially what it is about a person that gives his words or thoughts whatever meanings they have in the first place.
Metaphysical semantic theories attempt to deal with questions such as
semantic holism (q.v.). and whether meaning can be specified independently of a
thinker's environment (see Content: Wide and Narrow). Examples of such theories
are causal theories of reference, teleological (q.v.) theories
that try to explain meaning in terms of evolution, and informational
(q.v.) theories that construe meaning as some sort of covariation relation
between contentful states of the person and the phenomena in the world that
those states pick out. Linguistic and metaphysical semantic theories are largely
independent. For example, the difference between teleological and informational
theories has no impact on any issue in linguistic semantics.
CRS is in the
domain of metaphysical semantics: it says that the nature of meaning is
functional. It does not have anything very informative to say about linguistic
issues about particular languages or about how a language user works out the
meanings of sentences on the basis of the meanings of their component words. But
if correct, it can contribute to these enterprises by discouraging false and
confused foundational views. For example, Fodor once argued that almost all
concepts are innate on the ground that to learn a word, (1) one must learn its
definition; but, he said, (2) there are very few definitions that are anything
other than translations into another language of the same expressive power; so
(3) the language of thought must come equipped with translations of most English
words (e.g. 'zipper'). Many have found the conclusion hard to accept, for how
could evolution have provided us with innate concepts of meson and
zipper? Thus it is an advantage of CRS that it allows one to hobble this
argument by opening space for an alternative to (1): one can learn a word by
learning to use it, independently of any definition.
One major motivation for CRS is a functionalist (q.v.) approach to the mind
generally. Functionalism says that what makes a state a mental state is the role
it plays in interacting with other mental states in a creature's psychology.
This gives rise to a weak form of CRS: a state is meaningful (i.e.
hassome meaning or other) by virtue of the fact that it plays a certain
role in a person's psychology. This form offers a reply to theories that insist
a mind requires something more. for example, Searle (1980) has argued that
computers cannot understand language in virtue of their programs or, more
generally, by manipulating symbols in a certain way. He rests his case on a
thought experiment, the Chinese Room (q.v.), in which a monolingual English
speaker manipulates Chinese symbols by following rules that don't require him to
understand the meanings of the symbols he is manipulating. The rules are so
devised that he produces sensible responses in Chinese to any Chinese inputs.
Searle says that nonetheless he does not understand any Chinese‹he's just
mindlessly manipulating symbols. CRS motivates the "systems reply": If we can
program a computer to be intelligent, it won't be the central processing unit
(CPU) all by itself that is intelligent or that understand the symbols, but
rather all the complex relations between the CPU and other subsystems fo the
mind, e.g. for perception, reasoning and decision making. So the whole system
understands Chinese even if the person who is simulating the CPU does
not.
The weak form of CRS is functionalist about mental states (e.g.
beliefs vs. desires) without commitment to functionalism about mental
contents (believing that snow melts vs. believing that grass
grows). The former claims, in effect, that having any content at all
depends upon having a role in certain processes. The latter, along with CRS,
claims that the specific content a state has depends upon its role. The
functional role of a thought includes all sorts of causes and effects that are
non-semantic, e.g. perhaps happy thoughts can bolster one's immunity, promoting
good health. Conceptual roles are functional roles minus such non semantic
causes and effects.
What makes this stronger claim of CRS plausible is the
fact that many terms seem definable only in conjunction with one another, and
not individually in terms outside of the circle they form. For example, in
learning the theoretical terms of Newtonian mechanics ---'force', 'mass',
'kinetic energy', 'momentum', etc.-- we do not learn definitions outside the
circle. There are no such definitions. We learn the terms by learning how to use
them in our thought processes, especially in solving problems. Indeed, CRS
explains the fact, noted by Kuhn (1962), that modern scientists can't understand
the phlogiston theory without learning elements of an old language that
expresses the old concepts. The functional role of e.g. 'principle' as used by
phlogiston theorists is very different from the functional role of any term or
complex of terms of modern physics, and hence we must acquire some approximation
of the 18th Century functional roles if we want to understand their ideas.
Moreover, CRS does seem to give a plausible account of the meanings of the
logical connectives. For example, we could specify the meaning of 'and' by
noting that certain inferences --e.g., the inferences from p, q to 'p and q',
and the inference from 'p and q' to p-- have a special status (they are
"primitively compelling" in Peacocke's (1992) terminology).
A further
motivation for CRS is that it explains a reasonable version of a principle of
charity (one of a number of such principles to be found in the work of Quine and
Davidson) according to which we cannot rationally attribute irrationality to a
person without limit. Attributing unexplainable irrationality leads to a poor
match of roles. If the best translation yields poor enough matches, then the
alien conceptual system is not intelligible in ours.
Putnam (1975) raised what might seem to be a powerful objection to any CRS. He pointed out that many "natural kind concepts," such as water and gold, depend in part for their meaning upon something other than the role of a representation in a person's head, namely upon what happens to be in their external enviroment (see CONTENT: NARROW AND WIDE). Suppose, for example, there were a "Twin Earth" that was exactly like the earth except that, wherever the Earth has H2O, Twin Earth had a superficially similar, but chemically different substance, XYZ. Arguably, Twin Earthers who think, as they put it, 'Water is greenish' are thinking thoughts with different contents from their counterparts on Earth. Our thought is that water is greenish, theirs is that twin water is greenish. If this is true, then it would seem to present a difficulty for CRS, since the conceptual role of an Earthling's and her Twin Earth twin's inner representations would, ex hypothesi, be identical.
Some proponents of CRS have responded by favoring a "two-factor" version of CRS. On this view, meaning consists of an internal, "narrow" aspect of meaning--which is handled by functional roles that are within the body-- and an external referential/truth-theoretic aspect of meaning, which might handled by some of the other metaphysical theories of meaning (e.g. a causal one) that we mentioned earlier. According to the external factor, 'Superman flies' and 'Clark Kent flies' are semantically the same since Superman = Clark Kent; the internal factor is what distinguishes them. But the internal factor counts 'Water is more greenish than bluish' as semantically the same in my mouth as in the mouth of my twin on Twin Earth. In this case, it is the external factor that distinguishes them.
Two-factor theories gain some independent plausibility from the need of them to account for indexical thought and assertions (q.v.), assertions whose truth depends upon facts about when and where they were made and by whom. For example, suppose that you and I say 'I am ill'. One aspect of the meaning of 'I' is common to us, another aspect is different. What is the same is that our terms are both used according to the rule that they refer to the speaker; what is different is that the speakers are different. White (1982) generalized this distinction to apply to the internal and external factors for all referring expressions, not just indexicals.
In a two-factor account, the conceptual roles stop at the skin in sense and effector organs; they are "short-arm" roles. But CRS can also be held in a one factor version in which the conceptual roles reach out into the world--these roles are "long-arm". Harman (1987) has advocated a one factor account which includes in the long-arm roles much of the machinery that a two-factor theorist includes in the referential factor, but without any commitment to a separable narrow aspect of meaning. Harman's approach and the two-factory theory show that the general approach of CRS is actually compatible with metaphysical accounts of reference such as the causal theory or teleological theories, for they can be taken to be partial specifications of roles.
Actual conceptual roles involve errors, even dispositions to err. For
example, in applying the word 'dog' to candidate dogs, one will make errors, for
example in mistaking coyotes for dogs (see Fodor, 1987). This problem arises in
one form or another for all naturalistic theories of truth and reference, but in
the case of CRS it applies to erroneous inferences as well as to erroneous
applications of words to things. Among all the conceptual connections of a
symbol with other symbols, or (in the case of long-arm roles) with the world,
which ones are correct and which ones are errors? Kripke (1982), for example,
wonders what distinguishes someone who mistakenly says '57 + 65 = 5' from
someone who says it correctly, meaning by '+' a function that agrees with
addition except in yielding a value of 5 with 57 and 65 as arguments. The answer
a person gives in the two cases could be the same, correct in one and erroneous
in the other.
Some think we can solve the problem by appealing to
dispositions to "correct" their previous answers, or to "correct" those
corrections. But others wonder why all these dispositions could not be the same
for two persons who use '+' to designate different functions. (The problem of
error is sometimes said to be the problem of specifying semantic "norms",
although norms in this sense should not be confused with norms in the sense of
how one ought to apply a word; see Horwich, 1994.) Another line of reply is to
attempt to specify some sort of naturalistic idealization which specifies roles
that abstract away from error, in the way that laws of free fall abstract away
from friction.
Fodor has been a persistent critic of CRS, raising over the years a number of different objections to it. He criticizes a computer-oriented form of CRS for confusing what words denote with the words themselves. The functional roles in the target version of CRS stress searching data banks and manipulating representations, and this Fodor says is like claiming that the meaning of "Napoleon won at Waterloo" is a set of instructions for finding that sentence in a book in the New York Public Library. All such a search yields is more words, we never get the semantic values of those words, namely Napoleon or Waterloo.
From the point of view of a two-factor version of CRS, there are two mistakes in this criticism. First, according to CRS, the roles themselves are something over and above the words that have the roles; these roles are a kind of meaning or at least a component of meaning. Second, long-arm roles include causal chains outside the machine. And the two-factor version of CRS relies on a second factor, the referential factor, to explain the relation between the word 'Napoleon' and Napoleon.
CRS is often criticized from the point of view of truth-conditional theories of meaning. If the meaning of a sentence is its truth conditions, then the meaning cannot be its conceptual role. But with the two-factor theory, proponents of CRS have the option of counting meanings as the same or different in accord with whether the external factor specifies truth-conditions that are the same or different. Further, there is reason to suppose that meaning is more fine-grained than truth conditions. For example, the truth condition of 'I am happy' and 'Ned is happy' are the same, but the meanings of those sentences differ. The further machinery involved in the internal factor can capture the differences among sentences with the same truth conditions.
Fodor also criticizes CRS for giving the wrong account of how I and Helen Keller (who was blind and deaf from an early age) can mean the same thing by 'Water tastes great'. After all, none of her thoughts bear the same relation to the evidence of sight and sound that mine do. But here Fodor assumes that CRS only has the resource of appealing to similarity in inferential role, which is entirely internal. He disparages such an account in favor of a referential view: we mean the same because our concepts of water are concepts of the same thing. But a two-factor CRS, relying in part on a referential component, has the option of giving exactly the same account, as can a long-arm one factor account.
Fodor and LePore (1992) object to the two-factor account, wondering what glues the two-factors together. Why can't there be a sentence that has the inferential role of 'Water is greenish' but is true if and only if 3 is a prime number? But there is nothing in the CRS approach that dictates that there is any restriction at all on what roles can go with what truth conditions. This is an independent question that both proponents and opponents of CRS can ask. Everyone who accepts the existence of inferential roles and truth conditions should find the question meaningful, whether or not they think these are two-factors of meaning.
CRS is often viewed as essentially holistic,but the CRS theorist does have the option of regarding some proper subset of the functional roles in which an expression participates as the ones that constitute its meaning. Thus the subset could be taken to be those that are analytic; or as the primitively compelling inferences (Peacocke) plus those generated by them; or the explanatorily basic regularities (Horwich). One natural and common view of what distinguishes the meaning-constitutive roles is that they are analytic (q.v.). (An analytic sentence is one that is true by virtue of meaning alone, and we can regard an inference as analytic if it is valid by virtue of meaning alone.) Proponents of CRS are thus viewed as having to choose between accepting holism and accepting that the distinction between the analytic and synthetic is scientifically respectable, a claim that has been seriously challenged by Quine (1954/1976). Indeed, Fodor and Lepore (1992), argue that, lacking an analytic/synthetic distinction, CRS is committed to semantic holism (q.v.), regarding the meaning of any expression as depending on its inferential relations to every other expression in the language. This, they argue, amounts to the denial of a psychologically viable account of meaning.
Proponents of CRS can counters as follows: First, there is a question of whether a meaning-constitutive inference is thereby analytic. If what is meaning constitutive is analytic, then holistic versions of CRS need analyticity too, since they regard all inferences as meaning-constitutive. But if what is meaning constitutive is not thereby analytic, then neither holistic nor non-holistic versions of CRS need analyticity. So analyticity is not the issue between holistic and non holistic versions of CRS.
Second, proponents of CRS can reply that the view is not committed to regarding what is meaning constitutive as analytic. In terms of our earlier two factor account, they can, for example, regard the meaning-constitutive roles as those that are explanatorily basic in a narrow psychology -- they are the rules that explain other rules of use and determine narrow content. (See Horwich, 1994). Narrow content doesn't involve truth values -- these arise only with regard to wide content -- and so a fortiori it doesn't involve any commitment to truth by virtue of meaning alone.
A third approach to accommodating holism with a psychologically viable
account of meaning is to substute close enough similarity of meaning for strict
identity of meaning. That may be all we need for making sense of psychological
generalizations, interpersonal comparisons, and the processes of reasoning and
changing one's mind.
Fodor and LePore (1992) raise a further worry that links the metaphysical semantic issue with a linguistic one: a CRS would seem to risk violating compositionality, that is, the requirement that the meaning of a complex expression be a function (in the mathematical sense) of the meanings of its parts. It is widely thought that such a property of both language and thought is required to explain how human beings seem to able to grasp indefinitely many ever more complicated thoughts and how humans can learn to understand complex sentences on the basis of simple ones. CRS threatens this principle, since, Fodor and LePore say, the conceptual role of a complex non-idiomatic representation is not always a function of the conceptual roles of its parts. Someone who thinks that rattling snakes, especially, are dangerous is disposed to infer `This is dangerous' from `This is a rattling snake' for reasons that may not depend at all on any inferences they are disposed to make from `This is rattling' or `This is a snake' separately.
Advocates of non-holistic versions of CRS should regard the argument's assumption that all inferences are to be included in inferential roles as question begging.
Non-holistic versions of CRS can deal with compositionality by counting only
a subset of inferences as meaning constitutive. As mentioned above, these
inferences could be identified as the analytic ones, the explanatorily basic
ones, or as those that are primitively compelling or generated by them. The
threat to compositionality can be avoided by not counting the inference from
'This is a rattling snake' to 'This is dangerous' as part of the
meaning-constitutive roles of either sentence.
Advocates of holistic
versions of CRS may wish to go along with Fodor and LePore in assuming that all
inferences are part of inferential roles. They should point out that the
inferential role of 'rattling' and 'snake' is a matter not just of their roles
in isolation from one another, but also their roles in contexts involving
'rattling' and 'snake' together. The "rules of use" of these terms are
context-sensitive, not context-free.
Once we allow context-sensitive rules of use, compositionality can be
trivially satisfied. For example, we can characterize the meaning of a
word as an ordered pair, <X, Y>, where X is the set of inferences
to sentences containing the word and Y is the set of inferences
from sentences containing the word. This is a holistic version of the
view, for it includes the inference from 'rattling snake' to 'dangerous' in the
meaning of 'rattling' and 'snake', and this example goes proxy for the inclusion
of every inference in the meaning of every word involved in those
inferences.
Now the roles just mentioned satisfy the requirements of
compositionality from a metaphysical point of view without being a
psycholinguistic theory of the representations on the basis of which language is
learned or sentences are understood. It is no part of the task of the theorist
of metaphysical semantics to be a psycholinguist. Nonetheless, it should be said
that the natural psycholinguistic representations of words from a CRS point of
view would be ones that point towards the roles of sentences in which the word
appears.
CRS is more of a framework for a theory than an actual theory. There is no
agreement among proponents of this framework about how the roles are
constituted. By actual causal interactions among thoughts? All? Some? If some,
which ones? And what about systematically mistaken inferences (e.g. the
"gambler's fallacy")? Do widespread cognitive illusions contribute to the
determination of meaning? Or are the roles normative? If the roles are idealized
to avoid mistakes, how is the idealization supposed to be understood? Inference
can be understood in intentional terms or in purely causal terms, and the latter
would be preferable from the point of view of avoiding circularity in specifying
roles. And is there any way of distinguishing correcting an an old practice from
changing to a new one (Kripke, 1982)?
Many successful philosophical theories
are quite sketchy. Some say that CRS is no worse than many of them, but others
say that the problems in filling in these details involve difficulties that are
fatal to the whole project.
Wittgenstein, L. (1953) The Philosophical Investigations, Macmillan