Using parsed corpora for circumventing parsing

Abstract

Corpora tagged with part-of-speech and phrase structure information have been used for both exploratory data analysis as well as unsupervised learning of language models. These corpora have proved invaluable resources for research activities such as training part-of-speech taggers, disambiguating word-senses, detecting noun-phrases, inducing selectional restrictions, extracting argument structure and inducing probabilistic grammars.

In this paper, we present some new techniques that use parsed corpora, not for inducing grammars but for circumventing parsing as much as possible. In particular, we will describe how a parsed corpus using a wide-coverage Lexicalized Tree Adjoining Grammar (LTAG) is used for this purpose. The first technique exploits the fact that LTAGs represent dependency and constituency information in a uniform way. The second technique uses Explanation-Based Learning methodology to view parsing as Finite State Transduction. Both the techniques exploit the central notions of LTAGs — lexicalization, extended domain of locality and factoring of recursion from the domain over which dependencies are specified.

We would like to thank R. Chandrasekhar, Christine Doran, Mitch Marcus and Martha Palmer for their valuable comments.