Computational models of language universals: Expressiveness, learnability and consequences

Every linguist is struck by similarities among even the most different and most culturally isolated human languages. It is natural to assume that some of these common properties, these language universals, might reflect something about the way people can learn and use languages. In some relevant sense, some of these properties may arise and be maintained even in culturally isolated languages because of special restrictions on the range of structural options available for human language learners. A bolder idea is that some of these language universals may guarantee that the whole class of languages with such properties is ‘learnable’ in a relevant sense. While considerable progress has been made on finding ways to clearly articulate and assess possibilities of these sorts in precise computational models, there has also been a shift to more sophisticated versions of a long-standing traditional perspective: it may not be so much the formal structure of human languages, but the special kinds of fit between form and meaning that give human languages their most distinctive properties, in which case some early work on language acquisition may have characterized inappropriately difficult learning problems. A more reasonable perspective on the learners’ predicament may recognize a certain non-arbitrariness in the relation between structures and their semantic values, so that only certain kinds of structures are expected to carry certain sorts of semantic values. This can allow semantic properties of expressions to provide clues about syntactic structure, and vice versa, enriching the evidence available to the learner. This paper will review some fundamental results in this line of inquiry, from universals formulated in terms of expressive power of grammars, to results on learnable subsets of the languages defined by those grammars, leading finally to recent views on semantically-characterized grammatical universals. Even restricting attention to hypotheses that are most empirically secure and independent of any particular choice among the major traditions in grammatical theory, the modern perspective is surprising in many respects and quite different from anything that could have been conceived at the 1961 Conference on Language Universals (Greenberg, 1963).