An Efficient LR Parser Generator for Tree-Adjoining Grammars

Abstract

The first published LR algorithm for Tree Adjoining Grammars (TAGs [Joshi and Schabes, 1996]) was due to Schabes and Vijay-Shanker [1990] . Nederhof [1998] showed that it was incorrect (after [Kinyon, 1997]), and proposed a new one. Experimenting with his new algorithm over the XTAG English Grammar [XTAG Research Group, 1998] he concluded that LR parsing was inadequate for use with reasonably sized grammars because the size of the generated table was unmanageable. Also the degree of conflicts is too high. In this paper we discuss issues involved with LR parsing for TAGs and propose a new version of the algorithm that, by maintaining the degree of prediction while deferring the {``}subtree reduction{''}, dramatically reduces both the average number of conflicts per state and the size of the parser.