Evaluation of LTAG parsing with supertag compaction

One of the biggest concerns that has been raised over the feasibility of using large-scale LTAGs in NLP is the amount of redundancy within a grammar¿s elementary tree set. This has led to various proposals on how best to represent grammars in a way that makes them compact and easily maintained (Vijay-Shanker and Schabes, 1992; Becker, 1993; Becker, 1994; Evans, Gazdar and Weir, 1995; Candito, 1996). Unfortunately, while this work can help to make the storage of grammars more efficient, it does nothing to prevent the problem reappearing when the grammar is processed by a parser and the complete set of trees is reproduced. In this paper we are concerned with an approach that addresses this problem of computational redundancy in the trees, and evaluate its effectiveness.