Improving Parsing and PP Attachment Performance with Sense Information

Proceedings of ACL-08: HLT, pages 317–325, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Improving Parsing and PP attachment Performance with Sense Information Eneko Agirre IXA NLP Group University of the Basque Country Donostia, Basque Country e.agirre@ehu.es Timothy Baldwin LT Group, CSSE University of Melbourne Victoria 3010 Australia tim@csse.unimelb.edu.au David Martinez LT Group, CSSE University of Melbourne Victoria 3010 Australia davidm@csse.unimelb.edu.au Abstract To date, parsers have made limited use of se- mantic information, but there is evidence to suggest that semantic features can enhance parse disambiguation. This paper shows that semantic classes help to obtain significant im- provement in both parsing and PP attachment tasks. We devise a gold-standard sense- and parse tree-annotated dataset based on the in- tersection of the Penn Treebank and SemCor, and experiment with different approaches to both semantic representation and disambigua- tion. For the Bikel parser, we achieved a maximal error reduction rate over the base- line parser of 6.9% and 20.5%, for parsing and PP-attachment respectively, using an unsuper- vised WSD strategy. This demonstrates that word sense information can indeed enhance the performance of syntactic disambiguation. 1 Introduction Traditionally, parse disambiguation has relied on structural features extracted from syntactic parse trees, and made only limited use of semantic in- formation. There is both empirical evidence and linguistic intuition to indicate that semantic fea- tures can enhance parse disambiguation perfor- mance, however. For example, a number of different parsers have been shown to benefit from lexicalisa- tion, that is, the conditioning of structural features on the lexical head of the given constituent (Mager- man, 1995; Collins, 1996; Charniak, 1997; Char- niak, 2000; Collins, 2003). As an example of lexi- calisation, we may observe in our training data that knife often occurs as the manner adjunct of open in prepositional phrases headed by with (c.f. open with a knife), which would provide strong evidence for with (a) knife attaching to open and not box in open the box with a knife. It would not, however, pro- vide any insight into the correct attachment of with scissors in open the box with scissors, as the disam- biguation model would not be able to predict that knife and scissors are semantically similar and thus likely to have the same attachment preferences. In order to deal with this limitation, we propose to integrate directly the semantic classes of words into the process of training the parser. This is done by substituting the original words with semantic codes that reflect semantic classes. For example, in the above example we could substitute both knife and scissors with the semantic class TOOL, thus relating the training and test instances directly. We explore several models for semantic representation, based around WordNet (Fellbaum, 1998). Our approach to exploring the impact of lexical semantics on parsing performance is to take two state-of-the-art statistical treebank parsers and pre- process the inputs variously. This simple method allows us to incorporate semantic information into the parser without having to reimplement a full sta- tistical parser, and …

Content truncated to 500 words