WASP-Bench: a Lexicographic Tool Supporting Word Sense Disambiguation

WASP-Bench: a Lexicographic Tool Supporting Word Sense Disambiguation David Tugwell & Adam Kilgarriff ITRI, University of Brighton Lewes Road, Brighton BN2 4GJ, UK {David.Tugwell,Adam.Kilgarriff}©itri.bton.ac.uk Abstract We present WASP-Bench: a novel approach to Word Sense Disambiguation, also providing a semi-automatic environment for a lexicographer to compose dictionary entries based on corpus evidence. For WSD, involving lexicographers tackles the twin obstacles to high accuracy: paucity of training data and insufficiently ex- plicit dictionaries. For lexicographers, the com- putational environment fills the need for a cor- pus workbench which supports WSD. Results under simulated lexicographic use on the En- glish lexical-sample task show precision compa- rable with supervised systems1, without using the laboriously-prepared training data. 1 Introduction WASP-Bench2 is a web-based tool support- ing both corpus-based lexicography and Word Sense Disambiguation. The central premise be- hind the initiative is that deciding what the senses for a word are, and developing a WSD program for it, should be tightly coupled. In the course of the corpus analysis, the lexicographer explores the textual clues that indicate a word is being used in one sense or another; given an appropriate computational environment, these clues can be gathered and used to seed a boot- strapping WSD program. This strategy clearly requires human input for each word to be disambiguated, which may raise 1 It should be noted that the lower figure for recall reflects solely the fact that not all words were attempted due to time constraints. 2 The system has been developed under EP- SRC project M54971. A demo is available at http:jjwasps.itri.bton.ac.uk. The second author was also a co-ordinator for the SENSEVAL-2 evaluation exercise-to limit any conflict of interest only the first author was in- volved applying the system to the SENSEVAL-2 task and had no prior knowledge of the format of the task. ' 151 the objection that the lexicon is far too large for any word-by-word work to be viable. How- ever, the amount of human interaction needed is far less than that involved in preparing train- ing data3 and lexicographers are already in the position of having to inspect every word in the vocabulary. If they use a interactive tool such as the WASP-Bench to help them in this, then total coverage becomes a feasible proposition. 2 WASP-Bench Methodology The workbench is implemented in perl and uses cgi-scripts and a browser for user interaction. 2.1 Grammatical relations database The central resource is a collection of all gram- matical relations holding between words in the corpus. The corpus currently used in WASP- Bench is the British National Corpus4 (BNC): . Using finite-state techniques operating over part-of-speech tags, we process the whole cor- pus finding quintuples of the form: {Rei, Wl, W2, Prep, Position}, where Rei is a relation, Wl is the lemma of the word for which Rei holds, W2 is the lemma of the other open-class word involved, Prep is the preposition or parti- cle involved and Position is the position of Wl in the corpus. Relations may have null values for …

Content truncated to 500 words