Enhancing Performance of Lexicalised Grammars

Proceedings of ACL-08: HLT, pages 613–621, Columbus, Ohio, USA, June 2008. c⃝2008 Association for Computational Linguistics Enhancing Performance of Lexicalised Grammars Rebecca Dridan†, Valia Kordoni†, Jeremy Nicholson†‡ †Dept of Computational Linguistics, Saarland University and DFKI GmbH, Germany ‡Dept of Computer Science and Software Engineering and NICTA, University of Melbourne, Australia {rdrid,kordoni}@coli.uni-sb.de, jeremymn@csse.unimelb.edu.au Abstract This paper describes how external resources can be used to improve parser performance for heavily lexicalised grammars, looking at both robustness and efficiency. In terms of robust- ness, we try using different types of external data to increase lexical coverage, and find that simple POS tags have the most effect, increas- ing coverage on unseen data by up to 45%. We also show that filtering lexical items in a su- pertagging manner is very effective in increas- ing efficiency. Even using vanilla POS tags we achieve some efficiency gains, but when us- ing detailed lexical types as supertags we man- age to halve parsing time with minimal loss of coverage or precision. 1 Introduction Heavily lexicalised grammars have been used in ap- plications such as machine translation and informa- tion extraction because they can produce semantic structures which provide more information than less informed parsers. In particular, because of the struc- tural and semantic information attached to lexicon items, these grammars do well at describing com- plex relationships, like non-projectivity and center embedding. However, the cost of this additional in- formation sometimes makes deep parsers that use these grammars impractical. Firstly because, if the information is not available, the parsers may fail to produce an analysis, a failure of robustness. Sec- ondly, the effect of analysing the extra information can slow the parser down, causing efficiency prob- lems. This paper describes experiments aimed at improving parser performance in these two areas, by annotating the input given to one such deep parser, the PET parser (Callmeier, 2000), which uses lex- icalised grammars developed under the HPSG for- malism (Pollard and Sag, 1994). 2 Background In all heavily lexicalised formalisms, such as LTAG, CCG, LFG and HPSG, the lexicon plays a key role in parsing. But a lexicon can never hope to contain all words in open domain text, and so lexical cover- age is a central issue in boosting parser robustness. Some systems use heuristics based on numbers, cap- italisation and perhaps morphology to guess the cat- egory of the unknown word (van Noord and Mal- ouf, 2004), while others have focused on automati- cally expanding the lexicon (Baldwin, 2005; Hock- enmaier et al., 2002; O’Donovan et al., 2005). An- other method, described in Section 4, uses external resources such as part-of-speech (POS) tags to select generic lexical entries for out-of-vocabulary words. In all cases, we lose some of the depth of informa- tion the hand-crafted lexicon would provide, but an analysis is still produced, though possibly less than fully specified. The central position of these detailed lexicons causes problems, not only of robustness, but also of efficiency and ambiguity. Many words may have five, six or more …

Content truncated to 500 words