e-mail: kaye@umbsky.cc.umb.edu (internet)
[Note this paper includes an appendix. To access the appendix you may use the link above or one of the links available each time the appendix is mentioned in the footnotes to the text below]
Quine (1953) may be read as offering a challenge to those attempting to offer a positive account of meaning--and in particular, a naturalistic account: Provide a means of distinguishing those statements true in virtue of meaning alone-- analytic statements--from those whose truth depends on other matters such as facts about the world--synthetic statements. Until now, the challenge has not been met.
I will answer Quinean worries about the analytic/synthetic distinction, and about the concept of meaning, by developing a naturalistic, causal role theory of mental content. A naturalistic theory of content is motivated by the need to explain the nature of the representational states that are described by theories in cognitive psychology; specifically, this sort of account should specify physically respectable states of the world that constitute a mental state's having content (or meaning or intentionality). The type of naturalistic theory of content that I will focus on is a functional role theory(1) which accounts for the meaning of a mental state in terms of its relations to other mental states. Since it is an explanation of meaning, a functional role theory should confront Quine's concerns about meaning and analyticity; that is what I propose to tackle here.
My inquiry will begin with an examination of how the analytic/synthetic concerns arise, and how they are connected to an account of meaning--we shall see that they are, in fact, intimately connected to a functional role theory of meaning. I will then introduce and refine one variety of the functional role view--a causal role theory. We shall then consider the possibility of accepting Quine's abandonment of the analytic/synthetic distinction in favor of meaning holism; I shall argue that this is not a tenable option for a naturalistic account of mental content. Instead, I will suggest a means for sharpening the causal role account and show how this will allow us to draw an analytic/synthetic distinction for mental representations.
In order to understand Quine's attack on the analytic synthetic distinction,(2) and its relation to naturalistic and, in particular, to functional role theories of meaning, it useful to begin by noting the view that Quine was attacking, Carnap's (1947, 1950) view of meaning. Carnap proposed that meaning in a language could be reduced to a language's linguistic framework--the set of semantic rules that specifies the definitions and semantic interrelations of the terms in the language. According to this view, a language can be thought of as a set of syntactically specified sentences together with a set of rules which assign them meaning.
Carnap was, by his own admission, only directly concerned with artificial languages--he suggests no more than that the construction of such languages will allow us to explicate our actual (linguistic) concepts--where such explication might be revisionary. Yet it is natural to understand his account as an explanation of ordinary meaning. Quine's objection to applying Carnap's view to natural languages is, in effect, that natural languages don't wear their semantic rules on their sleeves. And, he argues, there is no other available means of separating semantic rules from the rest of our beliefs. If no such class of rules can be distinguished, then there is no reason to think that the analytic/synthetic distinction in Carnap's artificial languages reflects a similar distinction in natural languages.
Quine's attack is not just aimed at Carnap's view of analyticity and meaning. Any attempt to draw the analytic/synthetic distinction will fail, he claims, for there is no such distinction--the acceptance of any sentence depends both on the conventions of language users and on facts about the world:
The lore of our fathers is a fabric of sentences...It is a pale gray lore, black with fact and white with convention. But I have found no substantial reasons for concluding that there are any quite black threads in it, or any white ones. (Quine 1966, p. 132.)
Now, one might object that there is nothing easier than drawing the analytic/synthetic distinction: Analytic sentences are those whose truth or falsity is determined solely by their meanings, all others are synthetic. But this overlooks two important aspects of Quine's attack. First, there is the question of drawing the distinction in practice. To use Quine's example (1953, p. 32), we may ask, is "Everything green is extended" analytic or not? To be told that if it's analytic, then (if it's true) it's true solely in virtue of the meanings of these words is no help at all. So someone addressing Quine's attack needs to provide more than this definition of analyticity. Ideally, they should provide a decision procedure for sorting statements into the analytic and the synthetic.
This leads to the second, and more important set of points: First, it is the notion of analyticity that is under attack. After questioning the status of 'Everything green is extended', Quine continues:
Now does my indecision over this example really betray an incomplete understand, an incomplete grasp of the "meanings", of 'green' and 'extended'? I think not. The trouble is not with 'green' or 'extended' but with 'analytic'(1953), p. 32.
Moreover, the notion of meaning is under attack here as well. As the previous quote suggests, Quine is not challenging the fact that there are conventions among language users, but he is skeptical of the idea that these somehow produce distinct units of meaning in speakers (or in languages.) Rather, as he would portray it, meaning--i.e., the influence of convention--is spread out across the whole of the "web" of belief. On this conception, there is no principled difference between belief alteration (acceptance or rejection) that changes the meanings of sentences and belief alteration that brings ones beliefs into line with the experienced facts; semantic and epistemic influences on language and belief blend completely.
Quine's attack is(3) therefore a burden of proof challenge for someone wishing to retain any sort of atomistic view of meaning--i.e., for anyone who conceives of meanings as being somehow isolated from (other) effects on belief. It would seem that any such conception must allow that it is possible for there to be sentences whose truth is determined solely by meaning-- whether there are any such sentences in our languages or not. For if meaning is a separate and distinct factor in the contribution to truth-values of sentences, then even if all actual sentences' truth values, are, as the "thread quote" claims, a join result of meaning and other factors, it should be (e.g., logically) possible to have sentences whose truth-conditions are determined only by the meaning factor.(4) So, someone holding this conception should be able to specify an in-principle means of distinguishing analytic and synthetic statements. That is, an adequate account of meaning--whether naturalist or not--should explain what meaning is in such a way that would show us what, ideally, we should look for in order to determine whether a statement is analytic or synthetic. But, Quine claims, no such explanation is to be had-- all accounts of analyticity rely on the intensional notions of analyticity, meaning, synonymy, or necessity--they do not advance our ability to distinguish analytic and synthetic statements one bit, nor do they even point us in any helpful directions. So, the attack on the analytic/synthetic distinction amounts to an attack on a "discrete" notion of meaning, and, thus, on meaning itself.
It is important and useful to note the failure of two possible lines of support for a discrete notion of meaning and the analytic/synthetic distinction that are suggested in an influential reply to Quine by Grice and Strawson (1956). Their main argument in favor of an discrete conception of meaning is that the notions meaning, synonymy, analyticity, necessity, and the like form a set or "family" of interdefinable concepts that are quite useful, in spite of the fact that there appears to be no way of reducing or otherwise securing them to other (e.g., materialistic) notions. Now, first, this is not a line of reply that is open to a naturalist about meaning, for the idea of naturalism is to reduce or otherwise link concepts to workably empirical (e.g., scientific) concepts. But even the non-naturalist must confront the fact that Quine offers an alternative picture of what "meaning" (and the like) amount to--both in the web of belief conception of belief fixation, and in his subsequent claims that there is no more to meaning, synonymy, etc., than is revealed in translation (1960). The defender of this family of notions must, it would seem, provide an account of what meaning is that upholds these notions and provides a plausible alternative to Quine's explanations; otherwise, this line of reply is just dogmatic adherence to a certain philosophical conception and to certain common sense beliefs.
Grice and Strawson emphasize the fact that we have determinate intuitions about the analyticity of many sentences. Thus, one might propose that while there are some fuzzy boundary cases, such as Quine's "Everything green is extended", this does not show that a distinction cannot be drawn that decides the analytic/synthetic status of most sentences. Quine's line of reply to this suggestion seems to be--as the above quote suggests---to deny the obvious. Apparently, he would claim not to have determinate intuitions about the analyticity of any sentences. While this may not seem convincing, it does point out the problem with this reply of Grice and Strawson's, namely that any given intuition is subject to doubt. Thus, intuitions cannot provide a bedrock foundation for the notions of meaning and analyticity to rest on. Certainly, if it is supposed that knowledge of analyticity is a priori knowledge, then intuitions are inadequate to fulfill this role, since they are not immune to empirical skepticism such as Quine's. To use a Quinean phrase, intuitions alone do not provide a principled analytic/synthetic distinction; what is required is an explanation or account of meaning that will vindicate or justify the intuitions (and perhaps modify them).
So Quine's attack on the analytic/synthetic distinction stands as a burden of proof challenge to anyone attempting to explain and defend an atomistic notion meaning--this applies to the naturalist in particular. The challenge is, at bottom, to provide a principled means of distinguishing the semantic from the non-semantic contributions to our beliefs.
As noted at the outset, I intend to develop a naturalistic view of meaning that will meet Quine's challenge. My account is an instance of a general class of naturalistic views that may be called functional role theories of content for mental representations. The basic idea of a functional role theory of content is that the meaning of a mental state is a matter of the role it plays in the mental life of the cognizer. This view has obvious affinities to use theories of meaning--of the sort attributed to Wittgenstein and to Sellars.(5) According to these accounts, a word's meaning is constituted by how a community of speakers uses it. In like manner, for a functional role theory of content a mental symbol's content is constituted by the way it is used in the cognitive system.(6)
Of course, the notions of "functional role" and "use" are extremely general to the point of meaninglessness--it is the task of a plausible functional role theory to specify the sorts of conditions that are to count as the content--constituting role. The account I will present here selects causal relations as the relevant conditions.(7) It is standardly assumed that semantically evaluable mental states, e.g. beliefs, will also be states that have causal influences either on behavior or on other mental states. And theories in cognitive psychology attribute a rich set of causal processes to the mind/brain. So it is natural to look to causal roles in an account of representational content. Specifically, a causal role theory of content says that a mental state's (e.g., a mental term's) content is a matter of that state's causal role in cognition (i.e., the mind/brain).
However, the full causal account is much too broad. To see this we may note that a given state of a cognitive system will stand in countless causal relations that are clearly irrelevant to meaning. For example, suppose that in a given computer, information is stored in a certain location in memory, which amounts to there being a certain pattern of electrical charges present at a given location in the system. These charges will stand in countless causal relations that have nothing to do with their informational content. For instance, a bit of static electricity will alter them, and if a device testing for current flow is hooked up appropriately, they will (in part) cause a signal light to go on. But it is obvious that these sorts of causal relations are irrelevant to what gives these states an informational aspect--the obvious restriction one wants to make is that only the causal relations to machine (e.g., "program") operations in the computer that should be considered when assessing a state's informational content.
Similar considerations should apply to human mental states, which we may assume, are abstract types of neural firings. It is pretty clear that the causal relations of a mental state to headaches(8) or EEG probes aren't part of what constitutes the state's meaning, nor, it seems, should we think that all of its causal relations to other states at the level of atomic or subatomic particles are relevant to its content (and probably not even all of its biochemical causal relations are relevant.)
A related problem arises from the fact that it seems that patterns of causal relations alone are insufficient to manifest meaning. Consider the causal relations of a set of mental states specified in a completely abstract manner that provides no information about the intrinsic qualities of the states themselves. This is, according to the causal role theory, what constitutes the content of these states. But it is possible, and in fact even likely, that, even though these patterns may be highly complex, they can on occasion be found instantiated in the causal relations among subatomic particles--given the large number of highly intricate causal relations that are found between the subatomic particles of any bit of matter. But it is wildly implausible to think that every now and then, portions of the subatomic array become representational systems, e.g., instantiating the concept "horse" or the content "there's a blue book on the table."
So a simple causal role view appears inadequate because it fails to constrain the sorts of causal relations that will be allowed to count in the specification of content. What is needed is a means of selecting which types of causal relations will constitute meaning. The seemingly obvious idea is that a state's semantic causal relations constitute the state's content, but that, of course, is unacceptably circular for a naturalistic theory of content. In what follows, we shall try to find a naturalistic means of narrowing the causal relations down to the intuitively semantic.
A first major step towards this end involves the idea of limiting the allowable causal relations to those that are psychological. Of course, if "psychological" is defined in a way that makes essential use of semantic or intentional notions, then once again the account is circular. On the other hand, appeal may be made to psychological theories that cite semantic states as part of their explanations as long as the theories themselves are not selected in virtue of the fact that they describe the semantic. Toward this end, roughly, psychology may be defined as the science that attempts to explain the behaviors of complex organisms by characterizing the internal states that cause their behaviors.(9) A psychological role theory will identify contents of the internal states cited by this sort of theory as the causal roles of these states.
Alternatively, and preferably, a more fully developed scientific psychology may yield suitable states for a psychological role theory. Projecting the development of cognitive psychology, we may imagine a future set of (roughly computational) theories that, among other things, describe syntactic types of psychological states that constitute a natural kind. A psychological role theory can then appeal to the causal relations among these syntactically defined states. Note, again, that there is no problem with circularity if such a theory cites semantic as well as syntactic features of these states, as long as the semantic features are not used in the specification of causal relations. And this is the central tenet of computationalism--that all causal relations are syntactically specifiable. A psychological role theory will claim that the semantic properties are some subset of these syntactic causal roles.
A psychological role theory, then, identifies the meaning of psychological states with the psychological causal roles of those states--i.e. with those states causal relations to other psychological states, where "psychological" is defined in either of the ways just discussed. However, it may appear that this account is still not choice enough in its selection of causal states, since the psychological role view so far discussed effectively embraces Quine's holism about meaning because it allows that any causal relations of psychological states will count as part of the meaning of those states. Thus, any change in belief will constitute a (potential) change in meaning.
We have seen that Quine makes the point that there is no obvious means of distinguishing semantic from other, e.g., epistemic, influences on sentences. Since functional role theories identify a mental representation's meaning with its role in cognition, the lack of such a distinction would seems to lead immediately to Quine's alternative conception of system-wide meaning.(10) Specifically, for the causal role(11) account there appears to be no way of selecting those causal relations that are semantic. So all causal relations among psychological states are allowed to count in the determination of meaning.
It has been suggested, by Block (1991), that this is an acceptable consequence of a theory of content. The unrestricted causal role theory(12) amounts to (Quinean) meaning holism, the view that all (e.g., causal) cognitive liaisons contribute to terms and sentences meanings. However, this view appears to have very unpalatable consequences that we shall now consider.
The wide variety of implausible implications that may be cited against meaning holism tend to fall into two main categories. First, there are those objections that try to show that meaning holism undermines our ordinary conceptions of constancy of meaning in a given individual. For instance, it might be argued that if (psychological role) meaning holism is true, then we are never able to come to believe what we once believed to be false, or vice versa, since each change in belief results in a slightly different set of psychological states, hence a slightly different set of psychological roles, holisticly defined, hence a slightly different meaning for each mental state. Likewise, it might be claimed that meaning holism entails that there is no such thing as constructive reasoning, since each new conclusion that is accepted changes the contents of the states that were the believed premises. And, it would seem by similar considerations that there is no such thing as knowledge over time, since each change in mental states will yield a slightly different set of contents, so no set of contents remains constantly known through any changes in belief.
A second set of objections to meaning holism focuses on the fact that there appears to be a wide variance in beliefs, experiences and mental states between individuals. Certainly, there are never individuals that share the same set of mental states. But, the objection goes, if it is the total set of mental states that defines the content of mental states, then it will follow that no two individuals ever share exactly the same content--everyone will have at least slightly different content for their mental states, if not drastically different content. And this appears to be an absurd result. It means that there is no such thing as a shared set of concepts; indeed, it means that we never really communicate with one another, for the content that a given speaker intends to convey to her audience will as a rule not be something the audience can grasp, owning to differences in the total sets of mental states. Nor, it seems can their be such a thing as disagreement, if there is no common content on which to disagree. But all this seems patently absurd. Surely we at least occasionally communicate, disagree, and generally share concepts and content in our verbal interactions.
There are two response options available for the psychological role theorist here. The first is to simply accept these results and maintain that despite their apparent absurdity, they are acceptable consequences of a notion of content. The second option is damage control; the role theorist can attempt to block these implications or show that what follows along these lines are, in fact, only weaker threats to notions like reasoning and communication.
The first strategy does not seem very hopeful. It might be argued that we must be willing to accept some revisionary consequences of an otherwise acceptable theory of content. However, being asked to abandon our notions of reasoning, agreement and disagreement, and communication is extraordinarily revisionary. Surely, if there is any competing account of content--and there are several--which promises less revisionism, the alternative would be that much more appealing. Since I, for one, will provide a less revisionary account below, I take it that this line fails.
The second strategy appears, initially, to offer slightly more promise for the holistic psychological role theorist, particularly in regard to the problem of content differences between speakers. While it is clear that any two individuals will differ in many beliefs, e.g., most readers will differ in their beliefs about what the most plausible account of contents is and what its details are--still, it might be argued, there will be a tremendous amount of agreement as well--surely all readers will agree that dogs are animals, that 10+10=20 and that the Lincoln Memorial is in Washington, D.C. Thus, the role theorist might begin by maintaining that the objection overstates the differences and ignores the similarities among individuals--in a given community, the total variance in mental state sets may be much more modest than the objection claims, as consideration of the vast number of shared mundane beliefs shows.
Further, to rehearse an argument first made in a related regard by Katz(13), there is now substantial evidence from empirical psychology to support the view that many of our basic cognitive processes--in particular our perceptual processes--are largely innate. This is not to deny that each of us has differing experiences, but it does suggest that the possible range of those experiences may be fairly tightly constrained by our innately shared cognitive processes and equipment. If many aspects of cognition are innate, then this may provide important limits on the possible variance in the mental state sets of individuals.
From here, the (functional role) holist may attempt to minimize the apparent implausibility of holism by suggesting a modest revision in our conceptions of reasoning, communication, etc.--viz., that we replace our uses of the notion of constant or shared content with a notion of similar content. So, it might be suggested, someone has changed their beliefs when they believe the denial of something very similar in content to what they once believed. Or, it might be suggested, two people have communicated when the listener has come to believe that the speaker uttered something which is very similar in content to the speaker's actual utterance.
However, it is important to see that it will be extremely difficult to pursue this line of defense, for at least two reasons. First, it should be apparent that sometimes small differences in content do matter; imagine a valid argument with an enormously complex conjunctive conclusion. Another argument whose conclusion is the same except for the negation of the final conjunct apparently has a conclusion that is very similar in content, yet this slight change may, obviously, render the argument invalid. And philosophers are all too familiar with the fact that very slight differences in comprehension between speaker and audience can sometimes prove an extreme barrier to communication.
Second, note that the holist cannot legitimately cite the notion of shared beliefs in the argument rehearsed above, for beliefs are standardly individuated by content, but, the holist claims, content only comes from the total set of psychological states. Therefore, it must be something like assent to sentences that is shared. But, again, holism by its nature allows that the same sentences can acquire very different meanings in different psychological contexts. And everyone--even extreme nativists-- acknowledge that there is a wide variation in experiences between individuals. Everyone will likewise acknowledge that there are vast differences in sentences assented to between individuals and in individuals over time. The similarity found in mundane beliefs and innate constraints do not make these differences go away. As long as there are differences--whether they amount to 50%, 10% or 1% of psychological states--holism allows for differing contents. Thus, if holism is correct, it is not really apparent--despite shared assent of many sentences and shared innate constraints-- that we really share any beliefs at all.
Both these points show that if the holist is to make good on this line of defense, he must ultimately offer us some sort of system for measuring both inter-speaker and intra-speaker content similarity--a system that also tells us when small differences in content matter and when they do not. But this would seem to be an extraordinary difficult task--certainly, there is no such metric at present, nor even a clue as to how to go about developing one.(14)
Moreover, there is a further objection that will not prove susceptible to this sort of defense. The problem is that the holistic causal role view--and any functional role view that accepts holism--will imply that all sentence transitions, no matter how random, have at least some analytic warrant. Consider any sentence transition that occurs in a given speaker--where the affirmation of one sentence causes a second to be accepted too.(15) According the holistic functional role theory, analytic warrant will be conferred on all such sentence transitions, no matter how random. For example, if, as a result of random association, I am disposed to judge that: Eric Clapton is God, after judging that: The sky is blue, then the causal role theorist must admit that this is an analytic warranted inference, since the meaning of each of the terms for me is partly dependent on the fact that I accept this sentence transition, which is part of each's causal role. But it is patently absurd that any inferential warrant should be attached to a random association. Worse, any sentence transition, no matter how (otherwise) unwarranted will receive warrant as a result of causal role theory.
Perhaps the causal role theorist might consider accepting the latter consequence--so what if our random associations have more warrant--i.e., some--than we thought they did? But this is not really a very plausible option, since the assignment of analytic warrant to non-inferences threatens our notion of deductive validity. For an inference to be deductively valid is for it to have a logical form that is truth preserving. But, on the causal role view, any pair of true sentences will have a logical form that is truth preserving, i.e., the form that inferential role semantics uses to assign terms meaning, so any pair of true sentences that is transition-associated for a speaker will qualify as a deductively valid inference.(16) Surely, the contrast between valid and invalid arguments is quite a lot for the holist to ask us to give up.
Therefore, it appears that accepting a holistic causal role account of meaning threatens to undermine our notions of inference and validity, and this is surely not an acceptable consequence of a theory of meaning--or at least, it is a consequence to avoid at all costs. Fortunately, the atomistic account of meaning that I offer avoids holism, with all its implausibilities.(17)
I will now propose a means of limiting the psychological causal relations that will at the same time yield criteria for analyticity for mental representations. Consider mental sates, characterized syntactically. Which of those causal relations that a given psychological state stands in should intuitively matter to its meaning? I suggest that the best answer is those causal relations that it stands in owing solely to its structure as a symbolic state, or the structure of the system it exists in. This is most easily illustrated by considering the difference between association and predicate containment inferences. If an individual infers from A to B because the two (syntactic) representations are causally associated, then, it is not (solely) because of the form that either has, but rather because of the manner in which states of those types have occurred in cognition in the past--e.g., it may be because states of those types have frequently co-occurred. On the other hand, if an individual infers from 'x is AB' to 'x is A' it is (apparently)(18) because of the form--i.e., the predicate in the latter statement is contained in the complex predicate of the former statement. In the former case, the inference is drawn due to facts about the representations and their status in cognition that are completely independent of their structure, whereas in the latter case it is their make-up-how they are syntactically put together--that forms the basis for the inference.
The proposal, then, is this:
The causal relations that a representation stands in, in virtue of its structure as a representation and in virtue of the structure of cognition, are what constitute its meaning; all other causal relations are not part of its meaning.
Intuitively, the idea is this: the causal relations that depend on the constitution of representations and their causal relations to other elements of cognition that are part of the system's constitution are those causal relations that are necessary given the form of the representations and of the cognitive system--i.e., those causal relations such that, were they absent, that (syntactic) type of representation could not be present in that (structural) type of cognitive system.
For instance, it would seem that the justification of empirically based representations, e.g. percepts, is not a matter of the structure of the representations nor of the structure of cognition. Thus, if "A's are B's" is accepted based on inputs to the perceptual systems, then, nothing about the structure of 'A' or 'B' causes this acceptance and nothing about "A-detectors" or "B-detectors" in cognition cause any such representation to be accepted. By contrast, if "A's are B's" is accepted because of the meaning alone--i.e., it is analytic--then this should either be due to the fact that the make-up of "A"s and "B"s cause this acceptance or because the way that "A-detectors" and "B-detectors" are hooked-up in cognition causes all such representations to be accepted.
The notion of structure as it applies in the determination of content can also be nicely illustrated by the way that predicates receive their interpretations in computer languages. Take, for example, the fact that the string 'B0 00' as it occurs in the IBM 8088 system's machine language is interpreted as meaning "clear the accumulator" (or, more precisely, "move a '0' into the low portion of register A".) The reason it receives this interpretation is that because whenever the CPU reads that string, it will perform that operation. Note that the "meaning" assigned to the string is constant across locations where the string might be--i.e., it will be interpreted in this way regardless of whether it is found in a register, in memory or stored on a disk. Also note that this meaning is completely atomistic--any machine language program (for the 8088) containing this syntactic unit will be interpreted this way. Finally, the interpretation does not depend on the states the computer is in--it does not depend on what operation the computer has previously executed. The interpretation does depend on the fact that syntactic strings of this type, when read by the central processor, will cause the accumulator to be cleared. That is, it is this causal relation between the syntax of this string and the structure of the central processor, and nothing else, that determines the meaning (interpretation) of the string. If cognitive psychology is to make good on the computer metaphor, we should expect that future theories will characterize the content of mental representations in like manner, i.e., in terms of the causal relations that are due to the syntactic structure of representations and the relation of this structure to the structure of cognitive systems and processes.
What is the structure of cognition? Roughly, the architecture of the system. This may turn out be very dense or very sparse. I.e., nativists envision cognition as a set of rich systems, including a language faculty, a mathematics faculty, a music faculty, various perceptual modules and systems (e.g., vision), etc.(19) By contrast, radical empiricists envision cognition a little more than a massive associationistic system, whose only systematic features don't extend much beyond various principles of association and some activation and retrieval processes.(20)
What is not included in structure? As the example of the machine language element suggests, structural relations do not include relations of spatial and temporal location. Suppose that I now believe that socialism will arise from the decay of capitalism in virtue of a token of type 'socialism will arise from the decay of capitalism' being stored in my "belief box". Let us also suppose that such a token might have been stored elsewhere, say, in my "hope box", thereby instantiating the hope that socialism will arise from the decay of capitalism rather than the belief.(21) Intuitively, the token has the same meaning in either location. More generally, typical symbolically oriented views of cognition tend to assume that a given representation may be present with the same content in any of a number of different locations in cognition (e.g., storage, "on-line", in production) while retaining the same content. This is, in part, what the notion of systematic relations is intended to reflect. Moreover, we should expect this token to retain the same content regardless of what I have just been thinking or doing. That is, its content should not vary merely because of the context of other operations it occurs in. Thus, the structural relations are those that hold independent of where and when (vis-a-vis the occurrence of operations) the token occurs in cognition.
Now, it is easy to anticipate Quinean skepticism about this proposal. Surely, the Quinean will argue, there is no way of drawing a principled distinction between the causal relations in cognition that are structure-related and those that are not. This type of skepticism is, however, misplaced here.
Cognitive science already makes such distinctions in a limited form. For instance, there is relatively little difficultly in drawing the hardware/software distinction in computer science. Nor is there any difficulty in distinguishing a computer's capacities--those programs it can possibly run--from those it actually runs. And, specifically, we are readily able to distinguish the general causal role that a given computer-language predicate has vis-a-vis the architecture of the system from the causal role it has in an occurrence in a given program. It is not unreasonable to expect that computational cognitive theories--being developed within the computer paradigm--will allow for similar distinctions within the human mind/brain. The present problem is not that current cognitive theories fail to delineate the structural features of cognitive states; it is rather that such theories are still in the very early stages--we do not, as yet, have anything like a detailed map of what cognition is like. Indeed, the general lay of the terrain is still in doubt. Of course, skepticism about our ultimate ability to devise such theories is always possible. But that is different from "no principled distinction" concerns. Independent argument is need to show that ideal cognitive psychology can't possibly draw the structural/non-structural relations distinction.
Moreover, the structural, psychological role account will yield a criterion for analyticity, effectively meeting Quinean worries head-on. An analytic statement is one whose truth is determined solely by its meaning, therefore, an analytic statement is, on the proposed account, made true solely in virtue of the causal relations resulting from its syntactic structure and the causal relations that it stands in in virtue of the structure of the cognitive system that contains it.
But how can we establish which statements satisfy this characterization? The truth-determination of a representation will be reflected in a cognizer's reasons for accepting a given sentence.(22) A faculty of judgment, acceptance or belief fixation will reflect such reasons, and it is very likely that human cognition will contain at least one such faculty. To maintain consistency with naturalism, we must presume that such a faculty can be functionally (or otherwise non-semantically) specified. For instance, a faculty of judgment would appear to be one that approves representations for use in inference and (theoretical and practical) reasoning processes. If we have detected such a faculty, and have isolated the causes of the representations it accepts, then:
A representation, r, is an analytic truth iff it caused to be accepted solely in virtue of some combination of r's structure and the structure of the cognitive system (that contains r.)(23)
This specification can be developed in various ways as we learn more about the structure of our representations and the structure of cognition.(24) However, what we know already about the structure of natural languages(25) suggests the following types of analyticity for sentential representations:
(Category analyticity) r is analytic if r is a statement of the form 'A is B' and accepting 'x is A', for any 'x', ceteris paribus, causes it to be such that 'x is B' would be accepted if it was evaluated, solely in virtue of the fact that 'x is A' was accepted.
The idea is this. If 'rocks are minerals' is analytic (it may not be!) then accepting that something is a rock will cause acceptance of the thought that it is a mineral. Of course, accepting that something is a rock need not cause one to think the thought that it is a mineral; rather, one must be disposed to accept this thought if it occurs to one. Moreover, accepting that something is a rock may cause acceptance that it is a mineral if rocks and minerals are merely empirically associated in cognition. But in such a case, it is some non-structural fact, e.g., the conditioning of experience, together with the acceptance that x is a rock that leads to accepting x is a mineral. So it must be solely because of the fact that 'x is rock' is accepted that 'x is a mineral' is accepted if 'rocks are minerals' is to be analytic. That is, it must be because of the way that 'rock' and 'mineral' stand to one another and to the faulty of judgment that this judgment is accepted; it must not be true that this judgment might not have been accepted, if, say, the history of percepts received by the cognitive system had been different, even though the structure was the same.
Category analyticity will have the following special case:
(Analytic containment) r is analytic if r is a subject-predicate sentence and the predicate is contained in the subject.
This will apply to cases such as:
All A's are A's, e.g., All cows are cows
All A B's are B's, e.g., All brown cows are cows
These are perhaps the most obvious sort of analyticities, since it is readily apparent that the syntactic structure of the statements alone makes them true.
One other important class of analytic statements will be those that reflect the role of discrimination process in providing meaning for terms:
(Discriminatory analyticity) D is a content-assigning discriminatory ability for kind A if a positive output of x from D causes it to be such that 'x is D' would be accepted if it was evaluated, because of the positive output.
Sentences reflecting this sort of analyticity need not be typical of the language--they are apparently relatively rare in English (assuming that there are many such analytic connections)--but they will form an important part of a psychological theory that characterizes the semantics of representational states. Consider a dog detection unit (or routine) in a cognizer. Such a unit may have one of two sorts of connection to the cognizer's concept DOG; either it will be part of the meaning of DOG that everything satisfying the detection unit counts as a dog, or it won't. If not, the identification procedure is empirically connected to DOG, e.g., the cognizer may believe that dogs have the features satisfying the detector, but this is not an analytic belief. If so, then a positive output from the unit will be a sufficient condition for judging that the discriminated object is a dog. While it is unclear how many such analytic connections there are between our concepts and perceptual detection abilities, it is likely that there are a good number, since it appears that a fair amount of the content of our concepts rests on our perceptual knowledge of what sorts of things the world might contain.
This is only an initial sketch of types of analyticity, based on fairly robust intuitions about analyticity and corresponding facts about structure; it should be expandable to other classes of analytic statements. The present account of meaning also suggests an obvious criterion for the analyticity of inferences, viz.:
An inference is analytic iff accepting the premises causes it to be such that were the conclusion to be considered, it would be accepted solely in virtue of the structure of the premises (and of cognition.)
And, finally, the present account allows us to explain how there could be statements that we would affirm under any circumstances (e.g., "there are objects") even though they are, intuitively, non-analytic. According to the present account, such statements-- if there indeed are any--would not be affirmed for reasons having to do with the structure of the statement and cognition alone; rather, experience (or other features of cognition) would be part of the causes of the affirmation, even if the range of possible causes is so broad that any imaginable experience would allow for this affirmation. Thus, the present account of meaning promises an adequate explication of the notion of analyticity.
But what sort of answer to Quine is this? The proposed account of analyticity provides no actual help at all in deciding problematic cases, e.g., "Everything green is extended." Rather, it is a proposal concerning how to conceptually draw the distinction, with advanced empirical psychological knowledge that we do not now possess, along with some rough sketches suggesting how this means of deciding cases might support our intuitions about analyticity. How, then, does this meet Quinean skepticism about meaning?
It meets it by presenting a naturalistic (causal role) theory of content that is non-holistic, along with a naturalistic--and thus non-question begging--means of drawing the analytic/synthetic distinction. Together, the conception of meaning and the means of drawing the distinction meet Quinean worries about principle; the account of meaning underlies a means of, in principle, deciding the analytic/synthetic status of statements, and this distinction is founded on the conception of meaning--a naturalistic conception that takes us out of the circle of intensional notions. If this account is correct, Quine is wrong-- the problem with deciding borderline cases does not lie with the notions of analyticity and meaning, but rather it lies with our grossly incomplete knowledge of the causal relations within our mind/brains. The present view thus promises foundational support for philosophical and psychological uses of the notions of meaning, analyticity and the like.(26) And, though there is no guarantee, it may ultimately vindicate our intuitions about specific analyticities and meanings.
The present account thus promises support for the notions under attack by Quine, while writing off failures in our knowledge of meaning to our limited knowledge of empirical psychological facts. Notice the bold implication here: facts about content or meaning are not only not all knowable a priori--there is no reason to think that any of them must be a priori knowable. This raises very broad issues about meaning and our knowledge of it-- issues that I take up elsewhere.(27) Here, I will simply point out that a meaning naturalist should not expect there to be an a priori decidable analytic/synthetic distinction unless the naturalistic conditions that constitute meaning are knowable a priori. And those cited in the present account are obviously not the sort of conditions that we should expect to have a priori knowledge of.(28)(29) So, to the extent that Quine's attack is a challenge to the non-naturalist, a priori conception of meaning and semantics, the present account accords with Quine. However, unlike Quine, I am offering an alternative account that does not make meaning a priori knowable.(30)
Note that the present account, in rejecting holism, leaves open the question of whether a given predicate's or sentence's meaning depends on the presence of any other predicates or sentences. This account leaves it to empirical psychology to tell us if any of our mental representations' structure, or their relations to the structure of cognition require the presence of the representations--if so, then the representations' meaning will turn out to be "molecular", if not it will be "atomistic." But the general account does not dictate either atomism or molecularism for meaning--indeed, the safest bet is that cognition contains some representations with molecular content and some with atomistic content.
Finally, note that this may still leave us with some small dose of the worries that plague meaning holism. Specifically, although the present account is non-holistic in that it identifies meanings as independent sets of causal roles, nothing in the account lends support to the typical non-holist's assumption that communities of speakers share (most) meanings. I.e., this account does not--in virtue of what it says about content--entail that there is a single concept DOG that speakers either have or lack; it allows that possibility that the content attached to, e.g., 'dog' varies slightly or even substantially from speaker to speaker--itŐs all a question of how similar structural relations are across speakers. Now, it may be that the structure of representations and cognition are mostly innate, so perhaps there will be very little variance in content between speakers--but if this is so, it will again be empirical psychology rather than an analysis of meaning that verifies this constancy.
We have examined a naturalistic causal role theory of content that identifies the meanings of representations with the causal relations of representations that are determined by their structure and by cognition's structure. This account allows for a principled analytic/synthetic distinction--analytic (true) statements are those that are caused to be accepted solely in virtue of these structural relations. The view thus address Quine's concerns over the legitimacy of the notions of meaning and analyticity, while explaining why analytic and synthetic statements cannot be fully distinguished from the armchair (a priori.) The result is a very promising prospect for a theory of content.
Thanks to Ned Block, Nancy Daukas and Ernie Sosa for reading and commenting on this MS.
Block, N. (1986) "Advertisement for a Semantics for Psychology", in P. French, T. Uehling and H. Wettstein (eds.), _Midwest Studies in Philosophy, Vol. X_, University of Minnesota Press: Minneapolis, 615-678.
__________ (1987) "Functional Role and Truth Conditions", in _Proceedings of the Aristotelian Society_, Supplementary Vol. LXI, Blackwell: Oxford.
__________ (1991) "What Narrow Content is Not", in B. Loewer and G. Rey (eds.), _Fodor and His Critics_, Blackwell: Oxford, 33-64.
__________ (1993) "Holism, Hyper-analyticity and Hyper-compositionality", _Mind and Language_, 8:1, 1-26.
Carnap, R. (1947/1956) _Meaning and Necessity_ (2nd ed.) University of Chicago Press: Chicago.
__________ (1950) "Empiricism, Semantics and Ontology," reprinted in Carnap (1956).
Fodor, J. A. (1987) _Psychosemantics_. Cambridge, MA: MIT Press/Bradford.
Fodor, J.A. and E. LePore (1992a) _Holism: A Shopper's Guide._ Blackwell: Oxford.
__________ (1992b) "Why Meaning (Probably) Isn't Conceptual Role", _Mind and Language_, 6:4, 328-43.
Grice, H.P. and P.F. Strawson (1956) "In Defense of a Dogma", _Philosophical Review_, LXV, 141-58.
Harman, G. (1987) "(Nonsolipsistic) Conceptual Role Semantics", in _New Directions in Semantics_, E. LePore (ed.), Academic Press: London.
Jackendoff, R. (1987) _Consciousness and the Computational Mind_. MIT Press: Cambridge, MA.
Katz (1972) _Semantic Theory_. Harper & Row: New York.
Kaye, L.J. (1994) "The Computational Account of Belief", _Erkenntnis_ (forthcoming).
__________ (MSa) "Scientific Naturalism for Theories of Meaning"
__________ (MSb) "The Languages of Thought"
Putnam, H. (1983) "Computational Psychology and Interpretation Theory." In Putnam, _Realism and Reason, Philosophical Papers Vol. 3_, Cambridge University Press, pp. 139-154.
Quine, W. V. O. (1953), "Two Dogmas of Empiricism," in _From a Logical Point of View_. Harvard University Press: Cambridge, MA.
__________ (1966) "Carnap and Logical Truth", in his _The Ways of Paradox and Other Essays_, Harvard University Press: Cambridge, MA, 107-32.
__________ (1960) _Word and Object_. MIT Press: Cambridge, MA.