Learning a Complex Language From Simple Input: Contributions of Phrase Bracketing Information

210 p.Thesis (Ph.D.)--University of Illinois at Urbana-Champaign, 1983.All normal children succeed in acquiring the languages to which they are exposed. Yet, the rules which children induce are underdetermined by their language data. Moreover, children are not exposed to representative samples of their languages; rather, sentences addressed to children tend to be short and syntactically simple. Nevertheless, children exposed to different samples of a language succeed in inducing virtually identical sets of rules. How is the induction of such rules to be explained?This dissertation explores the ramifications for the problem of grammatical induction of the possibility that children's language input may include information directly revealing certain aspects of syntactic structure. In particular, such input may include cues which allow children to bracket sentences they hear into their constituent phrases. Such cues may include examples of individual phrases spoken in isolation (as is commonly found in speech addressed to children), prosodic cues marking phrase boundaries (such as phrase-final vowel lengthening), or distributions of function and content words (function words--articles, prepositions, etc.--typically occur in phrase-peripheral positions).A formal learnability-theoretic model of language acquisition is constructed, incorporating the possibility of input containing bracketing information as a basic assumption. This model rigorously demonstrates how children can induce complex transformational grammars on the basis of very simple input. A reasonable set of assumptions concerning the child's processing capacities are entailed by this model. In addition, this model allows for successful induction on a smaller quantity of data, and, hence, in a shorter amount of time, than have previous learnability-theoretic formulations. Thus, this is a feasible model of language acquisition.For a certain class of sentence constructions, this model predicts that learners should predominately arrive at particular analyses; previous models predict they should be equally divided across all possible analyses. A study examining acquisition of these constructions is presented; results indicate that these do, in fact, receive consistent analysis. This provides some corroboration of the bracketed input hypothesis.U of I OnlyRestricted to the U of I community idenfinitely during batch ingest of legacy ETD