Improved Default Sense Selection forWord Sense Disambiguation

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 11–17. Improved Default Sense Selection for Word Sense Disambiguation Tobias HAWKER and Matthew HONNIBAL Language Technology Research Group School of Information Technologies University of Sydney {toby,mhonn}@it.usyd.edu.au Abstract Supervised word sense disambiguation has proven incredibly difficult. Despite significant effort, there has been little suc- cess at using contextual features to accu- rately assign the sense of a word. Instead, few systems are able to outperform the de- fault sense baseline of selecting the high- est ranked WordNet sense. In this pa- per, we suggest that the situation is even worse than it might first appear: the high- est ranked WordNet sense is not even the best default sense classifier. We evalu- ate several default sense heuristics, using supersenses and SemCor frequencies to achieve significant improvements on the WordNet ranking strategy. 1 Introduction Word sense disambiguation is the task of select- ing the sense of a word intended in a usage con- text. This task has proven incredibly difficult: in the SENSEVAL 3 all words task, only five of the entered systems were able to use the contex- tual information to select word senses more ac- curately than simply selecting the sense listed as most likely in WordNet (Fellbaum, 1998). Suc- cessful WSD systems mostly fall back to the first- sense strategy unless the system was very confi- dent in over-ruling it (Hoste et al., 2001). Deciding which sense of a word is most likely, irrespective of its context, is therefore a crucial task for word sense disambiguation. The decision is complicated by the high cost (Chklovski and Mihalcea, 2002) and low inter-annotator agree- ment (Snyder and Palmer, 2004) of sense-tagged corpora. The high cost means that the corpora are small, and most words will have only a few ex- amples. The low inter-annotator agreement exac- erbates this problem, as the already small samples are thus also somewhat noisy. These difficulties mean that different sense frequency heuristics can significantly change the performance of a ‘base- line’ system. In this paper we discuss several such heuristics, and find that most outperform the com- monly used first-sense strategy, one by as much as 1.3%. The sense ranks in WordNet are derived from semantic concordance texts used in the construc- tion of the database. Most senses have explicit counts listed in the database, although sometimes the counts will be reported as 0. In these cases, the senses are presumably ranked by the lexicog- rapher’s intuition. Usually these counts are higher than the frequency of the sense in the SemCor sense-tagged corpus (Miller et al., 1993), although not always. This introduces the first alternative heuristic: using SemCor frequencies where avail- able, and using the WordNet sense ranking when there are no examples of the word in SemCor. We find that this heuristic performs significantly better than the first-sense strategy. Increasing attention is also being paid to coarse grained word senses, as it is becoming obvious that WordNet senses are too fine grained (Hovy et al., 2006). …

Content truncated to 500 words