Forest-guided Supertagger Training

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1281–1289, Beijing, August 2010 Forest-guided Supertagger Training Yao-zhong Zhang † Takuya Matsuzaki † † Department of Computer Science, University of Tokyo ‡ School of Computer Science, University of Manchester §National Centre for Text Mining {yaozhong.zhang, matuzaki, tsujii}@is.s.u-tokyo.ac.jp Jun’ichi Tsujii†‡§ Abstract Supertagging is an important technique for deep syntactic analysis. A super- tagger is usually trained independently of the parser using a sequence labeling method. This presents an inconsistent training objective between the supertagger and the parser. In this paper, we pro- pose a forest-guided supertagger training method to alleviate this problem by incor- porating global grammar constraints into the supertagging process using a CFG- filter. It also provides an approach to make the supertagger and the parser more tightly integrated. The experiment shows that using the forest-guided trained super- tagger, the parser got an absolute 0.68% improvement from baseline in F-score for predicate-argument relation recogni- tion accuracy and achieved a competi- tive result of 89.31% with a faster pars- ing speed, compared to a state-of-the-art HPSG parser. 1 Introduction Deep syntactic analysis by lexicalized grammar parsing, which provides linguistic-rich informa- tion for many NLP tasks, has recently received more and more attention from the NLP commu- nity. To use a deep parser in real large-scale ap- plications, speed is an important issue to take into consideration. Supertagging is one of the speed- up technique widely used for lexicalized grammar parsing. A supertagger is used to limit the number of plausible lexical entries fed to the parser, this can greatly reduce the search space for the parser. Supertagging was first proposed for Lexicalized Tree Adjoining Grammar (LTAG) (Bangalore and Joshi, 1999), and then successfully applied to Combinatory Categorial Grammar (CCG) (Clark, 2002) and Head-driven Phrase Structure Gram- mar (HPSG) (Ninomiya et al., 2006). In addi- tion, supertags can also be used for other NLP tasks besides parsing, such as semantic role label- ing (Chen and Rambow, 2003) and machine trans- lation (Birch et al., 2007; Hassan et al., 2007) to utilize syntactic information in the supertags. In lexicalized grammar parsing, supertagging is usually treated as a sequence labeling task inde- pendently trained from the parser. Previous re- search (Clark, 2002) showed that even a point- wise classifier not considering context edge fea- tures is effective when used as a supertagger. To make up for the insufficient accuracy as a single- tagger, more than one supertag prediction is re- served and the parser takes the burden of resolving the rest of the supertag ambiguities. A non-trivial problem raised by the separate training of the supertagger is that the prediction score provided by the supertagger might not be suitable for direct use in the parsing process, since a separately trained supertagger that does not take into account grammar constraints has a training objective which is inconsistent with the parser. Although the scores provided by the supertagger can be ignored (e.g., in some CCG parsers), this may also discard some useful information …

Content truncated to 500 words