Chapter 48 How There Could Be a Private Language and What It Must Be Like Jerry A. Fodor

The first objection I want to consider is an allegation of infinite regress. It can be dealt with quickly (but for a more extensive discussion, see the exchange between Harman, 1969, and Chomsky, 1969).

Someone might say: 'According to you, one cannot Ieam a language unless one already knows a language. But now consider that language, the metalanguage in which representations of the extensions of object language predicates are formulated. Surely, learning it must involve prior knowledge of a meta-metalanguage in which its truth definitions are couched. And so on ad infinitum. Which is unsatisfactory'. There is, I think, a short and decisive answer. My view is that you can't Ieam a language unless you already know one. It isn't that you can't Ieam a language unless you've already teamed one. The latter claim leads to infinite regress, but the former doesn't; not, at least by the route currently being explored. What the objection has in fact shown is that either my views are false or at least one of the languages one knows isn't learned. I don't find this dilemma embarrassing because the second option seems to me to be entirely plausible: the language of thought is known (e.g., is the medium for the computations underlying cognitive processes) but not learned. That is, it is innate. (Compare Atherton and Schwartz, 1974, which commits explicitly the bad argument just scouted.)

There is, however, another way of couching the infinite regress argument that is more subtle: ^zYou say that understanding a predicate involves representing the exten­sion of that predicate in some language you already understand. But now consider understanding the predicates of the metalanguage. Doesn't that presuppose a represen­tation of its truth conditions in some meta-metalanguage previously understood? And, once again, so on ad infinitum?' This argument differs from the first one in that the regress is run on 'understand' rather than on 1eam', and that difference counts.

What I said was that learning what a predicate means involved representing the extension of that predicate; not that understanding the predicate does. A sufficient condition for the latter might be just that one's use of the predicate is always in fact conformable to the truth rale. To see what's at issue here, consider the case of real computers.

Real computers characteristically use at least two different languages: an input/ output language in which they communicate with their environment and a machine language in which they talk to themselves (i.e., in which they run their computations). 'Compilers' mediate between the two languages in effect by specifying biconditionals whose left-hand side is a formula in the input/output code and whose right-hand side is a formula in the machine code. Such biconditionals are, to all intents and purposes, representations of truth conditions for formulae in the input/output language, and the ability of the machine to use that language depends on the availability of those definitions. (All this is highly idealized, but it's dose enough for present purposes.)¹ My point is that, though the machine must have a compiler if it is to use the input/output language, it doesn't also need a compiler for the machine language. What avoids an infinite regression of compilers is the fact that the machine is built to use the machine language. Roughly, the machine language differs from the input/output language in that its formulae correspond directly to computationally relevant physical states and operations of the machine: The physics of the machine thus guarantees that the se­quences of states and operations it runs through in the course of its computations respect the semantic constraints on formulae in its internal language.

I shall presently return to this point in some detail. For the moment, suffice it to suggest that there are two ways in which it can come about that a device (inducting, presumably, a person) understands a predicate. In one case, the device has and employs a representation of the extension of the predicate, where the representation is itself given in some language that the device understands. In the second case, the device is so constructed that its use of the predicate (e.g., in computations) comport with the con­ditions that such a representation would specify. I want to say that the first is true of predicates in the natural languages people Ieam and the second of predicates in the internal language in which they think.

3ut look', you might reply, 'you admit that there is at least one language whose predicates we understand without the internal representation of truth conditions. You admit that, for that language, the answer to: "How do we use its predicates correctly?" is that we just do; that we are just built that way. This saves you from infinite regress, but it suggests that even the regress from the natural language to the inner language is otiose. You argue that we Ieam "is a chair" only if we Ieam that it falls under the truth rule ^ryisa chair¹ is true iffx is G and then you say that the question of learning a truth rule for G doesn't arise. Why not stop a step sooner and save yourself trouble? Why not say that the question of how we Ieam "is a chair" doesn't arise either? Explanation has to stop somewhere'.

The answer is that explanation has to stop somewhere but it doesn't have to—and it better not—stop here. The question of how we Ieam 'is a chair' does arise precisely because English is learned. The question of how G is learned does not arise precisely because, by hypothesis, the language in which G is a formula is innate.

The critical property of the machine language of computers is that its formulae can be paired directly with the computationally relevant physical states of the machine in such fashion that the operations the machine performs respect the semantic constraints on formulae in the machine code. Token machine states are, in this sense, interpretable as tokens of the formulae. Such a correspondence can also be effected between physical states of the machine and formulae of the input/output code, but only by first compiling these formulae: i.e., only by first translating them into the machine language. This expresses the sense in which machines are %uilt to use' their machine language and are not %uilt to use' their input/output codes. It also suggests an empirical theory: When you find a device using a language it was not built to use (e.g., a language that it has learned), assume that the way it does it is by translating the formulae of that language into formulae which correspond directly to its computationally relevant physical states.

This would apply, in particular, to the formulae of the natural languages that speak- er/hearers learn, and the correlative assumption would be that the truth rules for predi­cates in the natural language function as part of the translation procedure.

Admittedly this is just a theory about what happens when someone understands a sentence in a language he has learned. But at least it is a theory, and one which makes understanding a sentence analogous to computational processes whose character we roughly comprehend. On this view, what happens when a person understands a sen­tence must be a translation process basically analogous to what happens when a ma­chine 'understands' (viz., compiles) a sentence in its programming language. I shall try to show that there are broadly empirical grounds for taking this sort of model seriously. My present point, however, is just that it is at least imaginable that there should be devices which need truth definitions for the languages they speak but not for the language that they compute in.

I don't, in short, think that the view of language learning so far sketched leads to infinite regress. It does lead to a one-stage regress; viz., from the natural language to the internal code—and that one stage is empirically rather than conceptually motivated. That is, we can imagine an organism which is bom speaking and bom speaking what­ever language its nervous system uses for computing. For such an organism, the ques­tion of how it Ieams its language would, ex hypothesi, not arise; and the view that its use of the language is controlled by an internal representation of the truth conditions upon the predicates of that language might well be otiose. All we would need to suppose is that the organism is so constructed that its use of the expressions in the language conforms to the conditions that a truth definition for the language would articulate. But we are not such organisms and, so far as I know, for us no alternative to the view that we Ieam rules which govern the semantic properties of the expressions in our language is tenable.

To begin with, it may be felt that I have been less than fair to the view that natural language is the language of thought. It will be recalled that the main objection to this view was simply that it cannot be true for those computational processes involved in the acquisition of natural language itself. But, though it might be admitted that the initial computations involved in first language learning cannot themselves be run in the language being learned, it could nevertheless still be claimed that, a foothold in the language having once been gained, the child then proceeds by extrapolating his boot­straps: The fragment of the language first internalized is itself somehow essentially employed to Ieam the part that's left.

Surely something that looks like this does sometimes happen. In the extreme case, one asks a dictionary about some word one doesn't understand, and the dictionary tells one, in one's own language, what the word means. That, at least, must count as using one part of one's language to Ieam another part. And if the adult can do it by the relatively explicit procedure of consulting a dictionary, why shouldn't the child do it by the relatively implicit procedure of consulting the corpus that adults produce? In partic­ular, why shouldn't he use his observations of how some term applies to confirm hypotheses about the extension of that term? And why should not these hypotheses be couched in a fragment of the very language that the child is learning; i.e., in that part of the language which has been mastered to date?

This begins to seem a dilemma. On the one hand, it sometimes does help, in learning a language, to use the language that one is trying to learn. But, on the other hand, the line of argument that I have been pursuing appears to show that it couldn't help. For I have been saying that one can't Ieam P unless one Ieams something like '^rP∕ is true iff Gx', and that one can't Ieam that unless one is able to use G. But suppose G is a predicate (not of the internal language but) in the same language that contains P. Then G must itself have been learned and, ex hypothesi, learning G must have involved learning (for some predicate or other) that G applies iff it applies. The point is that this new predicate must either be a part of the internal language or 'traceable back' to a predicate in the internal language by iterations of the present argument. In neither case however does any predicate which belongs to the same language as P play an essential role in media­ting the learning of P.

What makes the trouble is of course that the biconditional is transitive. Hence, if I can express the extension of G in terms of, say, H, and I can express the extension of P in terms of G, then I can express the extension of P in terms just of H (namely, ^ry is P^η) is true iff Hr. So, introducing G doesn't seem to have gained us any leverage. There doesn't seem to be any way in which the part of a natural language one knows could play an essential role in mediating the learning of the part of the language that one doesn't know. Paradox.

In fact, two closely related paradoxes. We want to make room for the possibility that there is some sense in which you can use one part of a language to Ieam other parts, and we want to make room for the possibility that there is some sense in which having a language might permit the thinking of thoughts one could not otherwise entertain. But the views we have so far been propounding seem not to admit of either possibility: Nothing can be expressed in a natural language that can't be expressed in the language of thought. For if something could, we couldn't Ieam the natural language formula that expresses it.²

Fortunately, both paradoxes are spurious and for essentially the same reasons. To begin with the learning case, what the argument thus far shows is this. Suppose F is a (proper) fragment of English such that a child has mastered F and only F at time t. Suppose that F' is the rest of English. Then the child can use the vocabulary and syntax of F to express the truth conditions for the predicates of F' only insofar as the semantic properties of F' teπns is already expressible in F. What the child cannot do, in short, is use the fragment of the language that he knows to increase the expressive power of the concepts at his disposal. But he may be able to use it for other purposes, and doing so may, in brute empirical fact, be essential to the mastery of F'. The most obvious possibility is to use F for mnemonic purposes.

It is a commonplace in psychology that mnemonic devices may be essential to a memory-restricted system in coping with learning tasks. If, as it seems reasonable to suppose, relatively simple natural language expressions are often coextensive only with quite elaborate formulae in the internal code, it becomes easy to see how learning one part of a natural language could be an essential precondition for learning the rest: The first-learned bits might serve to abbreviate complicated internal formulae, thus allowing the child to reduce the demands on computing memory implicit in projecting, confirm­ing, and storing hypotheses about the truth conditions on the Iater-Ieamed items. This sort of thing is familiar from teaching the vocabulary of formal systems. Complex

How There Could Be a Private Language 389 concepts are typically not introduced directly in terms of primitives, but rather by a series of interlinking definitions. The point of this practice is to set bounds on the complexity of the formulae that have to be coped with at any given stage in the learning process.³

Essentially similar considerations suggest how it might after all be the case that there are thoughts that only someone who speaks a language can think. True, for every predicate in the natural language it must be possible to express a coextensive predicate in the internal code. It does not follow that for every natural language predicate that can be entertained there is an entertainable predicate of the internal code. It is no news that single items in the vocabulary of a natural language may encode concepts of extreme sophistication and complexity. If terms of the natural language can become incorporat­ed into the computational system by something like a process of abbreviatory defini­tion, then it is quite conceivable that learning a natural language may increase the complexity of the thoughts that we can think. To believe this, it is only necessary to assume that the complexity of thinkable thoughts is determined (inter alia) by some mechanism whose capacities are sensitive to the form in which the thoughts are couch­ed. As we remarked above, memory mechanisms are quite plausibly supposed to have this property.

So, I am not committed to asserting that an articulate organism has no cognitive advantage over an inarticulate one. Nor, for that matter, is there any need to deny the Whorfian point that the kinds of concepts one has may be profoundly determined by the character of the natural language that one speaks. Just as it is necessary to distin­guish the concepts that can be expressed in the internal code from the concepts that can be entertained by a memory-restricted system that computes with the code, so, too, it is necessary to distinguish the concepts that can be entertained (salve the memory) from the ones that actually get employed. This latter class is obviously sensitive to the particular experiences of the code user, and there is no principled reason why the experiences involved in learning a natural language should not have a specially deep effect in determining how the resources of the inner language are exploited.⁴

What, then, is being denied? Roughly, that one can Ieam a language whose expres­sive power is greater than that of a language that one already knows. Less roughly, that one can Ieam a language whose predicates express extensions not expressible by those of a previously available representational system. Still less roughly, that one can Ieam a language whose predicates express extensions not expressible by predicates of the representational system whose employment mediates the learning.

Now, while this is all compatible with there being a computational advantage asso­ciated with knowing a natural language, it is incompatible with this advantage being, as it were, principled. If what I have been saying is true, than all such computational advantages—all the fadlitatory effects of language upon thought—will have to be explained away by reference to 'performance' parameters like memory, fixation of atten­tion, etc. Another way to put this is: If an angel is a device with infinite memory and omnipresent attention—a device for which the performance/competence distinction is vacuous—then, on my view, there's no point in angels learning Latin; the conceptual system available to them by virtue of having done so can be no more powerful than the one they started out with.

It should now be clear why the fact that we can use part of a natural language to Ieam another part (e.g., by appealing to a monolingual dictionary) is no argument against the view that no one can Ieam a language more powerful than some language he already knows. One cannot use the definition D to understand the word W unless (a) 'VV

means D' is true and (b) one understands D. But if (a) is satisfied, D and W must be at least coextensive, and so if (b) is true, someone who Ieams W by learning that it means D must already understand at least one formula coextensive with W, viz., the one that D is couched in. In short, learning a word can be learning what a dictionary definition says about it only for someone who understands the definition. So appeals to dictionaries do not, after all, show that you can use your mastery of a part of a natural language to Ieam expressions you could not otherwise have mastered. All they show is what we already know: Once one is able to express an extension, one is in a position to Ieam that W expresses that extension.

Notes

1. Someone might point out that, if the compiler formulae are biconditional they could be read as specifying truth conditions for formulae in the machine language with the input/output code providing the metalinguistic vehicles of representation. In fact, however, the appearance of symmetry is spurious even if the two languages are entirely Intertranslatable. For while the machine uses the machine code formulae without appealing to the compiler, it has no access to formulae in the input/output language except via the translations that the compiler effects. There is thus a useful sense in which, so far as the machine is concerned, machine language formulae express the meanings of formulae in the input/output code but not vice versa.

2. I know of only one place in the psychological literature where this issue has been raised. Bryant (1974) remarks: "the main trouble with the hypothesis that children begin to take in and use relations to help them solve problems because they Ieam the appropriate comparative terms like 'larger' is that it leaves unanswered the very awkward question of how they learned the meaning of these words in the first place." (p. 27) This argument generalizes, with a vengeance, to any proposal that the learning of a word is essential to mediate the learning of the concept that the word expresses.

3. I am assuming—as many psychologists do—that cognitive processes exploit at least two kinds of storage: a 'permanent memory' which permits relatively slow access to essentially unlimited amounts of information and a 'computing memory' which permits relatively fast access to at most a quite small number of items. Presumably, in the case of the latter system, the ability to display a certain body of information may depend critically on the form in which the information is coded. For extensive discus­sions see Neisser (1967). Suffice it to remark here that one way in which parts of a natural language might mediate further language learning is by providing the format for such encoding.

4. It should nevertheless be stressed that there is a fundamental disagreement between the kinds of views I have been proposing and those that linguistic relativists endorse. For such writers as Whorf, the psychological structure of the neonate is assumed to be diffuse and indeterminate. The fact about development that psychological theories are required to explain is thus the emergence of the adult's relatively orderly ontological commitments from the sensory chaos that is supposed to characterize the preverbal child's experience. This order has, to put it crudely, to come from somewhere, and the inventory of lexical and grammatical categories of whatever language the child Ieams would appear to be a reasonable candidate if a theorist is committed to the view that cognitive regularities must be reflexes of environmental regularities. On this account, the cognitive systems of adults ought to differ about as much as, and in about the ways that, the grammars and lexicons of their languages do and, so far as the theory is concerned, languages may differ without limit.

On the internal code story, however, all these assumptions are reversed. The child (indeed, the infraverbal organism of whatever spedes) is supposed to bring to the problem of organizing its experi­ences a complexly structured and endogenously determined representational system. Similarities of cognitive organization might thus be predicted even over wide ranges of environmental variation. In particular, the theorist is not committed to discovering environmental analogues to such structural biases as the adult ontology exhibits. He is thus prepared to be unsurprised by the prime fade inter­translatability of natural languages, the existence of linguistic universals, and the broad homologies between human and infrahuman psychology. (For further discussion, see Fodor et al., 1974.)

Bibliography

Atherton, M., and Sdiwartz, R., 1974. "Linguistic Innateness and Its Evidence." The Journal of Philosophy, 71:155-168.

Bryant, P. E., 1974. Perception and Understanding in Young Children. New Yorie Basic Books.

Chomsky, N_v 1969. "Linguistia and Philosophy." In S. Hook, ed. Language and Philosophy. New York: NYUPress.

Fodor, J_v 1975. The Language of Thought. Cambridge, MA: Harvard University Press.

Harman, G_v 1969. "Linguistic Competence and Empiricism." In S. Hook, ed. Language and Philosophy. New YoricNYUPress.

Neiser, U_v 1967. Cognitive Psychology. New Yorie Appleton.