Word Relatives in Context for Word Sense Disambiguation

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 42–50. Word Relatives in Context for Word Sense Disambiguation David Martinez Computer Science and Sofware Engineering University of Melbourne Victoria 3010 Australia davidm@csse.unimelb.edu.au Eneko Agirre IXA NLP Group Univ. of the Basque Country Donostia, Basque Country e.agirre@ehu.es Xinglong Wang School of Informatics University of Edinburgh EH8 9LW, Edinburgh, UK xwang@inf.ed.ac.uk Abstract The current situation for Word Sense Dis- ambiguation (WSD) is somewhat stuck due to lack of training data. We present in this paper a novel disambiguation al- gorithm that improves previous systems based on acquisition of examples by incor- porating local context information. With a basic configuration, our method is able to obtain state-of-the-art performance. We complemented this work by evaluating other well-known methods in the same dataset, and analysing the comparative re- sults per word. We observed that each algorithm performed better for different types of words, and each of them failed for some particular words. We proposed then a simple unsupervised voting scheme that improved significantly over single sys- tems, achieving the best unsupervised per- formance on both the Senseval 2 and Sen- seval 3 lexical sample datasets. 1 Introduction Word Sense Disambiguation (WSD) is an interme- diate task that potentially can benefit many other NLP systems, from machine translation to index- ing of biomedical texts. The goal of WSD is to ground the meaning of words in certain contexts into concepts as defined in some dictionary or lex- ical repository. Since 1998, the Senseval challenges have been serving as showcases for the state-of-the-art WSD systems. In each competition, Senseval has been growing in participants, labelling tasks, and tar- get languages. The most recent Senseval work- shop (Mihalcea et al., 2004) has again shown clear superiority in performance of supervised sys- tems, which rely on hand-tagged data, over other kinds of techniques (knowledge-based and un- supervised). However, supervised systems use large amounts of accurately sense-annotated data to yield good results, and such resources are very costly to produce and adapt for specific domains. This is the so-called knowledge acquisition bottle- neck, and it has to be tackled in order to produce technology that can be integrated in real applica- tions. The challenge is to make systems that dis- ambiguate all the words in the context, as opposed to techniques that work for a handful of words. As shown in the all-words tasks in Senseval- 3 (B. Snyder and M. Palmer, 2004), the current WSD techniques are only able to exceed the most frequent sense baseline by a small margin. We be- lieve the main reason for that is the lack of large amounts of training material for English words (not to mention words in other languages). Un- fortunately developing such resources is difficult and sometimes not feasible, which has been mo- tivating us to explore unsupervised techniques to open up the knowledge acquisition bottleneck in WSD. The unsupervised systems that we will apply on this paper require raw corpora and a thesaurus with relations between word senses …

Content truncated to 500 words