LAWS@SRI-AI.ARPA (12/17/84)
From: AIList Moderator Kenneth Laws <AIList-REQUEST@SRI-AI.ARPA> AIList Digest Sunday, 16 Dec 1984 Volume 2 : Issue 178 Today's Topics: Linguistics - Nonverbal Semantics ---------------------------------------------------------------------- Date: Fri, 14 Dec 84 13:13:04 EST From: Bruce Nevin <bnevin@BBNCCH.ARPA> Subject: Nonverbal Semantics [Long Message] >It's . . . convenient to talk about natural language as if >it's something "on its own". However, I view this attitude >as scientifically unhealthy, since it leads to an >overemphasis on linguistic structure. Surely the >interesting questions about NL concern those cognitive >processes involved in getting from NL to thoughts in memory >and back out again to language. These processes involve >forming models of what the speaker/listener knows, and >applying world knowledge and context. NL structure plays >only a small part in these overall processes, since the main >ones involve knowledge application, memory interactions, >memory search, inference, etc. Dyer V2 #160 >Bravo, Dyer! As you suggest, there is indeed much to learn >from the study of natural language -- but not about "natural >language itself"; we can learn what kinds of manipulations >and processes occur in the under-mind with enough frequency >and significance that it turns out to be useful to signify >them with surface language features. >. . . All that is very fine. We should indeed study >languages. But to "define" them is wrong. You define the >things YOU invent; you study the things that already exist. >. . . But when one confuses the two situations, as in the >subjects of generative linguistics or linguistic competence >-- ah, a mind is a terrible thing to waste, as today's >natural language puts it. Minsky V2 #162 I suspect that the antipathy to natural-language parsers, grammars, and theories that we often encounter in AI literature reflects a healthy revulsion from the excesses of generative linguistics. In all of its many schizmatic forms, generative grammer posits, as the secret inner mechanism of language, one of various language-like systems that share historical roots with programming languages, and uses natural-language data only in a fragmentary and anecdotal way to advance or refute the latest version. These systems can be quite hairy, but I am convinced that the hair is mostly inside the heads of the theorists. Any natural phenomenon, or any artifact of human culture, is a legitimate object of study. Natural language is both an artifact of human culture, and a natural phenomenon. There are some who are studying language, as opposed to the grammatical machinery of language-like systems. I recently reviewed a book by the linguist from whom Noam Chomsky learned about linguistic transformations (among other things). It will appear in AJCL vol. 10 nos. 3 and 4 (a double issue). The following excerpt gives an outline of the model of language he has developed: I refer to the Harrisian model of language as `constructive grammar' and to the Harrisian paradigm for linguistics as `constructive linguistics'. A constructive grammar has at least the following six characteristics: 1 The semantic primes are words in the language, a base vocabulary that is a proper subset of the vocabulary of the language as a whole. 2 Generation of sentences in the base is by word entry, beginning with entry of (mostly concrete) base nouns. The only condition for a word to enter is that its argument requirement must be met by some previously entering word or words, generally the last entry or entries, which must not already be in the argument of some other word. The base vocabulary has thus a few simple classes of words: N base nouns with null argument On, Onn operators requiring base nouns as arguments Oo, Ooo requiring operators as arguments Ono, Oon requiring combinations of operators and base nouns [NOTE: these are intended to be O with subscripts] This does not exhaust the base vocabulary. In addition to these, almost all of the operators require morphophonemic insertion of `argument indicators' such as -ing and that. (These were termed the `trace' of `incremental transformations' in Harris 1965 and 1968.) 3 The base generates a sublanguage which is informationally complete while containing no paraphrases. This is at the expense of redundancy and other stylistic awkwardness, so that utterances of any complexity in the base sublanguage are unlikely to be encountered in ordinary discourse. As in prior reports of H's work, base sentences are all assertions, other forms such as questions and imperatives being derived from underlying performatives I ask, I request, and the like. 4 A well-defined system of reductions yields the other sentences of the language as paraphrases of base sentences. The reductions were called the `paraphrastic transformations', and `extended morphophonemics' in earlier reports. They consist of permutation of words (movement), zeroing, and morphophonemic changes of phonological shape. Each reduction leaves a `trace' so that the underlying redundancies of the base sublanguage are recoverable. Linearization of the operator-argument dependencies--in English either `normal' SVO or a includes much of what is in the lexicon in generative `topicalizing' linear order--is accomplished by the reduction system, not the base. The reduction system includes much of what is in the lexicon in generative grammar (cf. Gross 1979). 5 Metalinguistic information required for many reductions, such as coreferentiality and lexical identity, is expressed within the language by conjoined metalanguage sentences, rather than by a separate grammatical mechanism such as subscripts. Similarly, `shared knowledge' contextual and pragmatic information is expressed by conjoined sentences (including ordinary dictionary definitions) that are zeroable because of their redundancy. [Harris's book Mathematical Structures of Language (Wiley 1968) shows that the metalanguage of natural language necessarily is contained within the language.] 6 The set of possible arguments for a given operator (or vice-versa) is graded as to acceptability. These gradings correspond with differences of meaning in the base sublanguage, and thence in the whole language. They diverge in detail from one sublanguage or subject-matter domain to another. Equivalently, the fuzzy set of `normal' cooccurrents for a given word differs from one such domain to another within the base sublanguage. In informal, intuitive terms, a constructive grammar generates sentences from the bottom up, beginning with word entry, whereas a generative grammar generates sentences from the top down, beginning with the abstract symbol S. The grammatical apparatus of constructive grammar (the rules together with their requirements and exceptions) is very simple and parsimonious. H's underlying structures, the rules for producing derived structures, and the structures to be assigned to surface sentences are all well defined. Consequently, H's argumentation about alternative ways of dealing with problematic examples has a welcome concreteness and specificity about it. In particular, one may directly assess the semantic wellformedness of base constructions and of each intermediate stage of derivation, as well as the sentences ultimately derived from them, because they are all sentences. By contrast, in generative argumentation, definitions of base structures and derived structures are always subject to controversy because the chief principle for controlling them is linguists' judgments of paraphrase relations among sentences derived from them. Even if one could claim to assess the semantic wellformedness of abstract underlying structures, these are typically so ill-defined as to compel us to rely almost totally on surface forms to choose among alternative adjustments to the base or to the system of rules for derivation. And as we all know, a seemingly minor tweak in the base or derivation rules can and usually does have major and largely unforeseen consequences for the surface forms generated by the grammar. This model of language offers an interesting approach to the problem brought up by Young in V2 #162, 174: how to represent the meaning of words without (circularly) using words? Most approaches amount to what I call `translation semantics': having found a set of language-universal semantic primes, one translates sentences of a given natural language into those primes and, voila'!, one has represented the `meaning' of those NL sentences. Let us ignore the difficulty of finding a set of semantic universals (a bit of hubris there, what!). The `representation of meaning' is itself a proposition in a more-or-less artificial language that has its own presumably very simple syntax (several varieties of logic are promoted as most suitable) and--yes--its own semantics. Logics boil `meaning' down to sets of `values' on propositions, such as true/false. `But my system', rejoins Young, `uses actual nonverbal modalities, it has real hooks into the neurological and cognitive processes that human beings use to understand and manipulate not only language, but all other experience as well'. That may be. It does beg the question to what degree cognitive processes and even neurological processes are molded by language and culture. (In Science 224:1325-1326 Nottebohm reports that the part of the forebrain of adult canaries responsible for singing becomes twice as large coincident with (a) increased testosterone and (b) learning of songs. This is the same whether the testosterone increase is annually in the Spring or experimentally, and the latter even in females, who consequently learn to sing songs as if they were males. Vrenson and Cardozo report in Brain Research 218:79-97 experiments indicating that both the size and shape of synapses in the visual cortices of adult rabbits changed as a result of visual training. Cotman and Nieto-Sampedro survey research on synapse growth and brain plasticity in adult animals in Annual Review of Psychology 33:371-401. Roger Walsh documents other research of this sort in his book Towards an Ecology of Brain. Conventional wisdom of brain science, that no new neurons are formed after infancy, is unwise.) The padres of yore surveyed the primitive languages around their missions and found so many degenerate forms of Latin. Their grammars lay these languages on the procrustean bed of inflections and declensions in a way that we see today as obviously ethnocentric and downright silly. We run the same risk today, because like those padres we cannot easily step out of the cultural/cognitive matrix with which we are analyzing and describing the world. Ask a fish to describe water: the result is a valid `insider's view', but of limited use to nonfish. Mr. Chomsky characterized his mentor in linguistics as an Empiricist and himself as a Rationalist, and in the Oedipal struggle which ensued mother Linguistics has got screwed. Given that systems based on constituent analysis are inherently overstructured, with layers of pseudo-hierarchy increasingly remote from the relatively concrete words and morphemes of language, an innate language-learning device is ineluctable: how else could a child learn all of that complexity in so short a time on so little and so defective evidence? The child cannot possibly be an Empiricist, she must be a Rationalist. Given a biologically innate language-acquisition device, there must be a set of linguistic universals that all children everywhere come into the world just knowing, and all languages must be specialized realizations of those archetypes--phenotypes of that genotype, as it were. (Chomsky did not set out to `define' natural language but to explain it. It is principally because his `underlying', `innate' constructs have a connection to empirical data that is remote at best--rather like the relation of a programmer's spec to compiled binary--that they appear to be (are?) definitions.) But consider a model in which the structure of language is actually quite simple. Might the characteristics of that model not turn out to be those of some general-purpose cognitive `module'? I believe Harris's model, sketched above, presents us this opportunity. Now about Jerry Fodor's book The Modularity of Mind, which Young mentions. The following is from the review by Matthei in Language 60.4:979, F presents a theory of the structure of the mind in which two kinds of functionally distinguishable faculties exist: `vertical' faculties (modules) and `horizontal' faculties (central processes). . . . F identifies the modules with the `input systems, whose function is to interpret information about the outside world and to make it available to the central cognitive processes. They include [five modules for] the perceptual systems . . . and [one for] language. . . . The central processes, as horizontal faculties, can be functionally distinguished from modular processes because their operations cross content domains. The paradigm example of their operation is the fixation of belief, as in determining the truth of a sentence. What one believes depends on an evaluation of what one has seen, heard, etc., in light of background information. . . . . . . the condition for successful science is that nature should have joints to carve it at: relatively simple subsystems which can be artificially isolated and which behave, in isolation, in something like the way that they behave in situ. (128) [The above, by the way, suggests that, while studying language in isolation--severing its `joints' with other systems--may be of limited interest to AI researchers seeking to model language users' performance, rather than their competence, it is not `scientifically unhealthy'. It also points to the central problem of semantics, as Matthei points out . . .] Modules, F says, satisfy this condition; central processes do not. If true, this is bad news for those who wish to study semantics. The burden which F puts on them is that they must demonstrate that computational formalisms exist which can overcome the problems he enumerates. These formalisms will have to be invented, because F maintains that no existing formalisms are capable of solving the problems. I, too, feel that notions of modules and modularity, or at least Fodor's attempt to consolidate them, make a great deal of sense. However, the caveat about the study of semantics underscores my contention that semantics properly must be based on an `acceptability model': a body of knowledge stated in sentences in the informationally complete sublanguage of Harris's base, whose acceptability is known. This is akin to a `truth model' in aletheic approaches to semantics in logic. It is also very simply conceived of as a database such as is constructed by Sager's LSP systems at NYU. We should note that the sentences of this base sublanguage correspond very closely across languages (cf. e.g. the English-Korean comparison in Harris's 1968 book), and that the vocabulary of the base sublanguage is a subset of that of the whole language (allowing for derivation, morphological degeneracy, and the like), much closer to Young's categories than the vocabulary with which he expresses so much frustration. There is one pointer I can give to another version of `translation semantics' that probably satisfies Young's sense of `nonverbal': Leonard Talmy developed an elaborate system for representing the semantics and morphological derivation of some pretty diverse languages in his (1974?) PhD dissertation at UC Berkeley. The languages included Atsugewi (neighbor and cousin to the Native American language I worked on), Spanish, and Yiddish. He went to SRI after graduation, but I have no idea where he is now or what he is doing. Bruce Nevin (bn@bbncch) ------------------------------ End of AIList Digest ********************