Abstract

Linguistic knowledge is on one view—the empiricist view—based primarily on linguistic experience: language acquisition proceeds by general purpose inductive mechanisms of analogy and generalisation. From this perspective the innate biases of the learner have little or no linguistically specific content. In this talk we will describe a precise mathematical theory of learning along these lines; based on what is called distributional learning. These learning algorithms can now learn classes of grammars that, on our current understanding are sufficiently powerful to describe all aspects of natural language syntax. We discuss three parts of the theory: the theory of weak learning, where we learn grammars that merely generate the right set of strings; strong learning, where we recover a grammar which generates an appropriate set of structural descriptions, and probabilistic learning, where we use indirect negative evidence to control overgeneralisation.

Taken together we will argue that these provide a plausible and well articulated theory of language acquisition based purely on distributional evidence, and without any problematic appeals to semantic bootstrapping.