How Can the Computational Approach Bridge the Two Disciplines, Brain Science and Linguistics?

Jun'ichi Tsujii 　

Abstract:
The ultimate goal of computational linguistics (CL) is to bridge the two scientific disciplines concerning language, brain science and linguistics. CL has attempted to construct computational accounts or processing models on how human brain treats language. In our view, such attempt can be divided into two, mutually related but somewhat independent, attempts. One is to develop formal descriptive frameworks for grammar and the other is their procedural realization. Although several competing descriptive frameworks such as HPSG, LTAG, CCG, etc., have been proposed, there is significant convergence among them. In particular, from the second point of view, i.e. the view point of procedural realization, most of them are mutually transferable.
   In this talk, we will present our work how to construct computationally feasible processing models from grammar in a formal framework, and how mutually translatable different formal grammars are. Single grammar can be partitioned into a set of successive grammars with different constraints. Furthermore, procedural grammars thus produced are combined with probabilistic models that are to be learned from actual language corpora, and thereby seemingly opportunistic human linguistic behaviour can be captured. In this talk, we try to put our work, which is intended to develop efficient NLP tools for practical engineering applications, in broader scientific context.