Language Learnability

The modern study of language acquisition began about twenty-five years ago, and it is no coincidence that its birthday followed fairly closely upon the birth of what we now call cognitive science.! One of the first human abilities that was shown to be "cognitive," in the sense of best being explained in terms of mental computations or rules acting on representations, was the child's acquisition of the syntax of his or her first language (Chomsky, 1959; Brown, 1958; Berko, 1958; Braine, 1963). It became clear early on that to understand why children say what they say, and how they learn to speak as adults, one must know what is in the child's mind, and that what is in the child's mind is best described in terms of "rules," "categories," "operations," and the like. The versions of functionalism, information-processing psychology, and cognitivism developed by Chomsky (1957), Newell and Simon (1961), Miller, Galanter, and Pribram (1960), Broadbent (1958), and Putnam (1960) sanctioned the mentalistic approach to language acquisition that was being forced upon developmental psycholinguists by their own empirical investigations.

At the same time, language acquisition appeared to be an ideal testing ground for rival theories about the precise nature of human computational mechanisms. Linguistic knowledge was recognized to be extremely intricate, yet children acquired it uniformly and effortlessly. It became clear that specific approaches to cognition would succeed or fail according to how well they could account for this feat. Thus in the early 1960s language acquisition came to occupy a central place in debates within psychology, linguistics, philosophy of mind, and-later-computer science.

For two reasons, however, the field of language acquisition has disappointed early hopes that a cognitivist theory of acquisition would emerge from a detailed study of children's language. First, as the connection between children's language and linguistic theory became more obscure (see Fodor, Bever, and Garrett, 1974; Pinker, 1982; Wexler and Culicover, 1980), developmental psycholinguists' attention turned away from using child language data to illuminate the process of language acquisition and turned toward characterizing children's language in its own terms, without regard to how the child ever attains adult linguistic proficiency (witness the popular use of the term "child language" for the field that used to be called "language acquisition").

The most exciting promise of the early field of developmental psycholinguistics was that it would provide an account of the mental processes by which a child interprets ambient speech and thereby induces the rules of the language of the adult community. Now that this goal has largely been abandoned, debates about the nature of children's word combinations assume less cross-disciplinary theoretical interest.

The second disappointment of developmental psycholinguistics has been its inability to arrive at a consensus as to how to solve its less ambitious goal, that of characterizing children's language. The catalog of computational machinery from which psycholinguistic theories could draw has proved to be an embarrassment of riches. The problem was no longer one of finding a single coherent model to account for the child's abilities but of deciding among alternative accounts, anyone of which could handle the data as well as the others. It turned out to be nearly impossible to determine which grammar best fit a corpus of child speech (Brown, Fraser, and Bellugi, 1964; Brown, 1973), especially since the rules in the grammars proposed often contained arbitrary mixtures of adult syntactic categories, ad hoc syntactic categories, semantic features, phonological features, and specific words (see Atkinson, 1982, for discussion). Furthermore, as the field of developmental psycholinguistics matured, more and more classes of cognitive mechanisms, with far greater flexibility than grammatical formalisms alone, became available to the theorist. Comprehension strategies (Bever, 1970), for example, can be stated in terms of any information that the child's mind can entertain, from linear order and syntactic category membership to knowledge about the probabilities of physical events. Since the advent of the "method of rich interpretation," whereby the investigator uses the context of an utterance to posit linguistic structures that are not revealed in the utterance itself, models of the mental representations underlying children's abilities have also included symbols for general thematic relations such as "agent" and "location" (e.g., Brown, 1973); specific semantic relations such as "eater" and "eaten entity" (Bowerman, 1973); and speech acts and pragmatic intentions (e.g., Halliday, 1975; Bruner, 1975). Fine-grained analyses of corpora of early speech have led others to suggest that the child uses rote patterns (R. Clark, 1971; MacWhinney, 1982) or utterance templates containing variables and constants (e.g., Braine, 1976), with the constant terms allowed to consist of specific words, semantic classes of various degrees of inclusivity (as in "custody for the purpose of oral ingestion"), or classes defined by features (MacWhinney, 1982). The child is also given credit for a variety of formal devices acting in concert with his or her rule systems, such as deletion operators (Bloom, 1970), output order biases (MacWhinney, 1982), elementary transformational operations (Erreich, Valian, and Winzemer, 1980), attentional filters (Slobin, 1973; Newport, Gleitman, and Gleitman, 1977), and rules of interpretation and construal (Lust, 1981). Indeed, some investigators, (e.g., MacWhinney, 1982) have come to the conclusion that to understand child language, one must posit a large set of relatively heterogeneous mechanisms