A Simple Approach for HPSG Supertagging Using Dependency Information

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 645–648, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics A Simple Approach for HPSG Supertagging Using Dependency Information Yao-zhong Zhang † Takuya Matsuzaki † † Department of Computer Science, University of Tokyo ‡ School of Computer Science, University of Manchester §National Centre for Text Mining, UK {yaozhong.zhang, matuzaki, tsujii}@is.s.u-tokyo.ac.jp Jun’ichi Tsujii†‡§ Abstract In a supertagging task, sequence labeling models are commonly used. But their lim- ited ability to model long-distance informa- tion presents a bottleneck to make further im- provements. In this paper, we modeled this long-distance information in dependency for- malism and integrated it into the process of HPSG supertagging. The experiments showed that the dependency information is very in- formative for supertag disambiguation. We also evaluated the improved supertagger in the HPSG parser. 1 Introduction Supertagging is a widely used speed-up technique for lexicalized grammar parsing. It was first proposed for lexicalized tree adjoining grammar (LTAG) (Bangalore and Joshi, 1999), then extended to combinatory categorial grammar (CCG) (Clark, 2002) and head-driven phrase structure grammar (HPSG) (Ninomiya et al., 2006). For deep parsing, supertagging is an important preprocessor: an ac- curate supertagger greatly reduces search space of a parser. Not limited to parsing, supertags can be used for NP chunking (Shen and Joshi, 2003), se- mantic role labeling (Chen and Rambow, 2003) and machine translation (Birch et al., 2007; Hassan et al., 2007) to explore rich syntactic information con- tained in them. Generally speaking, supertags are lexical tem- plates extracted from a grammar. These templates encode possible syntactic behavior of a word. Al- though the number of supertags is far larger than the 45 POS tags defined in Penn Treebank, sequence la- beling techniques are still effective for supertagging. Previous research (Clark, 2002) showed that a POS sequence is very informative for supertagging, and some extent of local syntactic information can be captured by the context of surrounding words and POS tags. However, since the context window length is limited for the computational cost reasons, there are still long-range dependencies which are not easily captured in sequential models (Zhang et al., 2009). In practice, the multi-tagging technique pro- posed by Clark (2002) assigned more than one su- pertag to each word and let the ambiguous supertags be selected by the parser. As for other NLP applica- tions which use supertags, resolving more supertag ambiguities in supertagging stage is preferred. With this consideration, we focus on supertagging and aim to make it as accurate as possible. In this paper, we incorporated long-distance in- formation into supertagging. First, we used depen- dency parser formalism to model long-distance re- lationships between the input words, which is hard to model in sequence labeling models. Then, we combined the dependency information with local context in a simple point-wise model. The experi- ments showed that dependency information is very informative for supertagging and we got a compet- itive 93.70% on supertagging accuracy (fed golden POS). In addition, …

Content truncated to 500 words