Statistical learning of syntactic structure: Mechanisms and constraints

To what extent can humans learn linguistic structure from statistical information in language input? This dissertation investigates the role of statistical learning in the acquisition of one aspect of syntax: phrase and hierarchical structure. In particular, the studies asked whether learners could organize serially-presented linguistic input into a hierarchically-organized phrase structure grammar by detecting the statistical properties of languages which mirror the dependencies, or predictive relationships, that obtain between form classes. A related hypothesis pursued in this research is that human statistical learning mechanisms are constrained in a fashion which focuses the learner's analyses on exactly the kinds of statistics pertinent to linguistic structure.;Six experiments investigated the learning abilities of adult and child subjects in laboratory tasks in which subjects were exposed to artificial languages. These languages were designed to probe the use of statistical information in language acquisition. The results suggest that human learners possess learning mechanisms which detect and utilize statistical cues signaling phrase and hierarchical structure. Moreover, these mechanisms are constrained to use those cues which are most characteristic of human languages. Learners imposed predictive dependencies even when the input, unlike natural languages, did not contain predictive cues to phrase structure. Under these circumstances, subjects changed the language to more closely resemble the structure of natural languages. Taken together, the results of these studies suggest an important role for constrained statistical learning in the acquisition of language.