The Science of Language

by Noam Chomsky

  One part of the explanation of the attractions of externalist semantics lies in incorrectly tying what seems to make language meaningful (its use) to the task of constructing a theory of linguistic meaning. This appears to suggest that the subject matter of a theory of meaning for natural languages be their uses. Instead, a theory's subject matter should be the internal cognitive capacities and systems that make the use of language in meaningful ways possible. Pragmatists such as Dewey helped create the illusion that the subject matter is use itself. They noted that it is in its use in preparing for and carrying out actions and projects that language becomes meaningful, intending by this that its significance for surviving and thriving becomes obvious. Note that if this were true of language, it should also be true of any form of cognitive capacity we have, if it were true at all. Our minds provide us with several cognitive resources, such as those found in the way it configures visual objects (cf. Biederman 1987). Add to this its ways of configuring tastes, sounds, human faces, actions, etc. Any of these cognitive resources can be meaningful or significant in this way too, for each is used by all of us to configure experience and understand the world and other persons and organisms, in order to survive and thrive. But no one sensible tries to offer a naturalistic theory of how the mind can help configure visual objects by making the subject matters of their theories a catalogue of how people use internally constituted resources. The sensible approach tries to say what the mind's cognitive resources are directly.

  Perhaps another part of the explanation lies in the fact that internalist semantics is not just at an early stage of development, but that the way proposed for constructing such a science conflicts with some preconceived views of how to construct a naturalistic science of mind. Consider, for example, Patricia Churchland's (1986, 2002) view that one must look directly to the brain to construct a science of the mind. The internalist approach to linguistic meanings cannot currently look to neurons, axons, and neural firing rates. That is because unless one has a theory in hand of what neural systems ‘do’ – of the computations they carry out – looking directly at neurons is as sensible as groping in the dark, even though that is what some do, or at least propose. Moreover, there is no guarantee at all that current understandings of neural systems and how they operate are on the right track. Just as physics had to change in order to accommodate to chemistry (in order to deal with bonding), it is quite possible that study of neural operations will have to change to accommodate the best computational theories of parts of the mind. So there is no good reason to take seriously the idea that a computational theory of language's meanings with its internalist and nativist assumptions be rejected because it has not met someone's view that the only science of mind worth taking seriously now is one that begins with some current understanding of neural operations and tries to explain (typically by appeal to training) how to impose on what are usually supposed to be ‘plastic’ neural systems a specific set of operations. It is likely much more fruitful to construct a theory and then look for ways to ‘embody’ its operations.

  To speak to both of the two potential blocks to looking inside the head outlined so far, internalists construct – postulate – modular and natively endowed computational systems in the head that take specific inputs, subject them to specified algorithms/operations, and yield specified domains of outputs. Their theories offer functions in the formal, mathematical sense that link elements to others and specify what the linked elements are. A computational theory of language and its meanings is a theory of this sort. To oversimplify, it takes lexical items that associate ‘sound features’ (phonological features) and ‘meaning features’ (semantic features), and specifies the combinatorial system that can combine them to yield at two interfaces with other systems (perceptual/production and conceptual/intentional) packaged phonetic and semantic ‘information’ that constitute the language faculty's contribution to the overall operations of the mind/brain. Marr-inspired computational theories of vision differ not in the basic assumption concerning how to proceed (in the methodology that can now best be employed in constructing a science of mind), but – unsurprisingly – in specified ‘inputs’ (arrayed input intensities), algorithms (‘Mexican hat’ mathematical formulae, among others), and ‘outputs’ (‘color’-‘position’, ‘textural’, etc. arrays). The ‘outputs’ are the ways that the visual system helps configure/plays a role in configuring the experience of the organisms that have such a system. Thought of from the different, non-scientific point of view of ourselves as experiencers and agents, rather than organisms, one can make sense of how the visual system, language, and so on help explain how we (as persons) conceive/perceive/sense/understand the world and others, and – especially with the contribution of language – in how we think. Continuing from within this point of view, no doubt we use the information that native internal systems provide us in order to survive and thrive; and this is part of what makes language and the other systems we have meaningful in Dewey's sense. But the naturalistic theory of a system is not a theory of how we use its output, or of why its contributions are meaningful to us. It is a theory of the ‘information’ it provides other internal systems, information that – returning to the personal point of view – we use. These points are as correct for language as for any other cognitive system.

  In sum, a naturalistic theory of meanings for a language is a theory of the ‘tools,’ ‘concepts,’ ‘perspectives’ that language offers the organism. It is a theory of what they – these concepts: individually rich, collectively complex and structured – become when gathered systematically by the combinatorial system in expressions/sentences. It is a theory of what they are, of how they develop in individuals and the species, and of how they are put together. So again: if our use of linguistic tools/meanings in certain ways happens to make them appear particularly meaningful or significant, it does not tell us what these tools are. So one can grant that – in certain respects – language seems especially meaningful or significant to humans (in effect, its utility becomes obvious) where it aids our efforts to cooperate with others in order to successfully prosecute various projects aimed at solving practical problems, from shopping to deciding how to climb a mountain with others to figuring out how one is going to vote. But a theory of meaning for language, like a theory of the ‘tools’ that vision offers, is a naturalistic theory of an internal mental organ, not of how it is used, or why. These natively provided and endless numbers of meanings, offered to us in organized forms in the expressions of our languages, are needed in order to solve practical problems. The right way to look at the matter – the internalist points out – is to take the existence of these native resources as necessary conditions of understanding and prosecuting actions in the way humans do, and to focus one's ‘theory of meaning’ on them – on their natures, on how they compose, and on how they develop in the individual human and in the species.

  Perhaps still another part of the explanation of the attractions of externalism lies in the commonsense concept LANGUAGE. From the naïve to the relatively sophisticated, if you ask someone what they think a language is, they are very likely to tell you – although not in these exact words – that it is a human invention, something like a rule-governed institution that people over millennia have put together in order to understand and deal with the world. And they will tell you that they learned language and came to participate in this supposed public institution because of training on the part of their parents and, more generally, the linguistic community in which they grew up. In some cases, they will believe this so strongly that they will deny obvious facts about how children do acquire language and conceptual resources and, if they are academics but remain wedded to the commonsense conception of language, they will devote a lot of time, effort, and research resources to trying to convince others that that conception of language must be right. We look at some such cases in the next subsection.

  Finally, probably another part of the explanation – at least for those who are sophisticated en
ough – lies in the apparent success in mathematics of Fregean semantics, exemplified in Frege's Begriffschrift (1879/1976) and in a more informal way in his “Sense and Reference” (1952). Many have employed Frege's techniques in efforts to construct semantic theories for natural languages. But there is a problem with these efforts. It is revealed in the first few pages of “Sense and Reference.” There Frege states his assumptions concerning the ‘languages’ to which his semantic efforts apply. He says that these ‘languages’ (really, symbol systems) are those found in specific communities of users, such that for each member of the community and each such ‘language,’ there is a determinate and unambiguous relationship of each “name” to a single “sense” of the name, and for each such sense, there is again a determinate relationship to a specific “denotation.” The reverse is not correct: a specific referent can be the target of multiple senses, hence ‘names.’ He illustrates by saying that two different ‘names’ (“the morning star” and “the evening star”) have different senses, even though the two senses have a single denotation (the planet Venus). I use scare-quotes around the term ‘language’ because Frege's primary interest was in constructing an externalist semantics for mathematics, and with the possible exception of the natural numbers and elementary mathematics, it is very unlikely that mathematics has much in common with natural languages. Indeed, Frege agreed: natural languages are too much of a mess; later philosophers, such as Tarski, agreed too. In a way Quine did too: he insisted on ‘regimenting’ natural languages, attempting to turn them into something that they are not, and – I emphasize – cannot be. Returning to Frege: for one thing, mathematics in its advanced forms is invented, not a free gift of native resources. For a second, especially relevant one, mathematicians choose to be careful in their use of their ‘languages,’ and typically are so. Working mathematicians are concerned with proof and provability and so do their best to use the technical terms of their fields to refer unambiguously to a single item or class of items. That is why it is possible to ‘find’ determinate referential/denotational relationships in mathematics, and (a similar case) in specific natural sciences. That is because the relationships are made to be determinate: the community makes them so. The relationships are ‘established’ and have their determinate-seeming character by virtue of the efforts of mathematicians. One can even argue that the ‘objects’ of the mathematical enterprise – aleph-null, for example – are ‘made’ by the efforts of mathematicians. In contrast, it is hard to find users of natural language who – apart perhaps from circumstances where they too are engaged on a project where precision and explicitness are crucial – have any intention or need to strictly restrain the uses to which the terms of their natural languages are put. For some discussion, see Chomsky (1995a, 1996).

  As mentioned, Frege himself recognized the limitations of his semantic efforts. While he illustrated his views of sense and denotation/reference with some natural language examples, he explicitly acknowledged that natural languages are not used in the way the symbols of mathematics are. This can be explained by pointing to the fact that the determinate mapping from name to sense to referent that his semantics relies on is found only where people are faithful in their efforts to use their terms in the same way. These efforts are found in mathematics and the practices of natural scientists who invent theoretical terms and apply them in what they try to make determinate ways. So in no case – whether in the use of the words of natural languages, or the invented symbols of mathematics and natural science – is a sign–thing relationship fixed in nature. ‘Fixing’ depends on the intentions and actions of the users of the relevant words and symbols.

  Nevertheless, there have been various attempts to accommodate Fregean semantics to natural language use. Some have adapted the idea that there is a uniform word–sense or word–denotation relationship of the sort Frege demanded to different contexts of use by introducing indices to deal with context; the result is an attempt to make word(s)-in-contexts that people produce, not just words, period, into the initial elements of the Fregean word–sense–denotation/referent chain. It is an attempt to circumvent one of the reasons Frege himself gave when he warned readers against employing his semantics with natural language use. However, there are serious problems with it. Perhaps indexing works with times of utterance and individual who utters – assuming that there is some agreed-on way to index these. But it is hard to conceive of how to deal in a determinate way with contextual factors such as ‘the topic of current discourse.’ No doubt we humans, relying on familiarity and our natural cognitive resources, make reasonably good judgments in our discussions and conversations about what a topic of discourse is; we rely on similar I-languages, familiarity with another person, story backgrounds, and so on. In doing so, however, we depend not on some kind of agreed-upon scheme of indexation and mappings-in-contexts, but on whatever resources and background knowledge we can muster for the occasion. It may sometimes appear automatic, and even seem determinate. But it is not something that any science has been able to deal with; assigning a value to the index “current topic of discussion” is completely post hoc. No doubt one can propose that there is some system in the head that accomplishes this feat. So far as I can tell, however, the ‘system’ introduced amounts to something like a homunculus with the remarkable capacity to do what we do when we undertake to understand – probably only approximately, but often sufficiently adequately to succeed at carrying out whatever task is at stake – what another means or intends. That would not be sufficient in mathematics or theorizing within a natural science domain.

  Efforts to turn Fregean semantics into a semantics for natural languages must confront other problems too. One must perform some sleight of hand to deal with the fact that fictional terms such as “Pegasus” and descriptions such as “the average Irishman” and “the square circle” make perfectly good sense to ourselves and to others when we speak and engage in conversation, even though the ‘world’ seems to lack average Irishmen and a winged horse named Pegasus. And square circles are particularly daunting. No one has difficulty using or understanding The square circle horrifies geometricians, even though square circles are ‘impossible objects.’ And then there is the problem of vagueness: bald (as in Harry is bald) does not have a determinate denotation/extension. And so on. I will not catalogue or discuss efforts to deal with these issues. No matter what restrictions, qualifications, and oddments of theoretical machinery are introduced, no general theory is on the horizon and, far more fundamentally, nothing provides a serious answer to (nor plausibly can speak to) what appears to be a fact, that reference is a matter of human use of terms and sentences, and that use appears to be free, and not written into natural events in a way that allows us to construct a theory. While Frege's views about how to proceed in doing semantics make some sense of the practices of mathematicians, he was right to say that they have limited application to natural languages or – more correctly – to the use of natural languages. And since the semantics he offers for mathematics depends essentially on the cooperation of groups of mathematicians in how they use their terms, his ‘semantic theory’ is not, and cannot be, a naturalistic science of meaning. The internalist demands nothing less than that.

  But what about Fodor's attempt to construct an externalist (albeit partially nativist) theory of meaning for natural languages, a theory that purports to have senses that determine their denotations/referents, only in a ‘naturalistic’ way? Fodor, recall, has a view of what he calls “concepts” and semantic relationships to the world that according to him is based on causal principles and purports to be a naturalistic theory of natural language meaning. It focuses on his view that meanings must be public (similarity in concepts apparently will not do), and his assumption that in order for them to be public, meanings must be identified with their ‘wide contents,’ where these are taken to be properties of things ‘out there.’ This does not rule out entirely a contribution of the mind: the concepts that natural languages express are
claimed to have a mental component too. Intuitively, Fodor holds that a concept is in part a mental entity consisting of a “mode of presentation” (a psychological/mental version of a Fregean sense) and a denotation, with the latter serving as the meaning of a natural language's term. And these concepts are claimed to be the topics of naturalistic theorizing. To explain: things ‘out there’ via the operations of causal impingements on the human sensory system bring about the acquisition of an internal representation in the form of a mode of presentation (MOP), which is what I have been construing as a lexically represented concept. So far, there is little to dispute; of course internal MOPs (or in the terminology above, the ways in which internal systems configure experience and understanding) develop or grow because of ‘input,’ and the input is informational in a good technical sense: the probability that a child/organism will develop a DOGmop from some impingements with doggish characteristics D is greater than from impingements with cattish ‘shapes’ C, where it is more probable that it will acquire a CATmop. What count as doggish characteristics? Fodor is not entirely clear about the matter, but on a reasonable interpretation, being doggish depends not on dogs, but on the nature of the internal MOP-production system and what it demands for specific ‘triggering’ inputs. Fodor assumes as much by saying that concepts such as DOG and all the others that figure in common sense are “appearance properties.”2 The problem lies instead in believing, as Fodor seems to, that the mode of presentation in turn stands in a semantic-denotational relation to the thing(s) that serve as the distal cause of the relevant impingements. In effect, he claims that a MOP m caused by some distal entity or property stands in a denotational relationship to that entity or property, and represents the causing thing(s)/properties ‘out there,’ and that this denotation is somehow determined by the nature of the MOP and the human mind. In his (1998) Concepts, one finds very little on how this determination is supposed to take place: he says only that we have the kinds of minds that “generalize to” a specific denotation-external property. In his (2008) LOT 2, he tries to expand on that mysterious capacity of the mind by introducing the idea that the mind is so assembled that some specific MOP ‘falls into’ the right denotation/external content. The picture he draws is at best very speculative. Assuming MOP m is associated with the term “m,” these external properties/things constitute the meaning of the term “m” (a syntactic entity) that is linked to the MOP. In this way, the linguistic term is ‘about’ its denotation.