Abstract

The fundamental problem of linguistics is to find how knowledge of language is represented and how that knowledge is acquired by children learning their first language; understanding or solving this problem would open the door to a new generation of intelligent language processing systems. This is fundamentally a computational problem, which can be studied using the tools of formal language theory and computational learning. Solving it requires reconceptualising some basic concepts—including the relationship between a grammar and the language it defines.

In this talk I will give an overview of this field (assuming no prior knowledge of linguistics or machine learning) and discuss some recent technical results in distributional learning that can potentially provide a solution to this problem. These techniques involve modelling the relationship between substrings and the contexts that they can appear in—these give rise to algorithms for learning classes of context free and context sensitive languages that seem to be a good match for the properties of natural language.