Unsupervised WSD based on Automatically Retrieved Examples: The Importance of Bias

Unsupervised WSD based on automatically retrieved examples: The importance of bias Eneko Agirre IXA NLP Group University of the Basque Country Donostia, Spain eneko@si.ehu.es David Martinez IXA NLP Group University of the Basque Country Donostia, Spain davidm@si.ehu.es Abstract This paper explores the large-scale acquisition of sense-tagged examples for Word Sense Disam- biguation (WSD). We have applied the “WordNet monosemous relatives” method to construct auto- matically a web corpus that we have used to train disambiguation systems. The corpus-building pro- cess has highlighted important factors, such as the distribution of senses (bias). The corpus has been used to train WSD algorithms that include super- vised methods (combining automatic and manually- tagged examples), minimally supervised (requiring sense bias information from hand-tagged corpora), and fully unsupervised. These methods were tested on the Senseval-2 lexical sample test set, and com- pared successfully to other systems with minimum or no supervision. 1 Introduction The results of recent WSD exercises, e.g. Senseval- 21 (Edmonds and Cotton, 2001) show clearly that WSD methods based on hand-tagged examples are the ones performing best. However, the main draw- back for supervised WSD is the knowledge acqui- sition bottleneck: the systems need large amounts of costly hand-tagged data. The situation is more dramatic for lesser studied languages. In order to overcome this problem, different research lines have been explored: automatic acquisition of training ex- amples (Mihalcea, 2002), bootstrapping techniques (Yarowsky, 1995), or active learning (Argamon- Engelson and Dagan, 1999). In this work, we have focused on the automatic acquisition of examples. When supervised systems have no specific train- ing examples for a target word, they need to rely on publicly available all-words sense-tagged corpora like Semcor (Miller et al., 1993), which is tagged with WordNet word senses. The systems perform- ing best in the English all-words task in Senseval-2 were basically supervised systems trained on Sem- cor. Unfortunately, for most of the words, this cor- 1http://www.senseval.org. pus only provides a handful of tagged examples. In fact, only a few systems could overcome the Most Frequent Sense (MFS) baseline, which would tag each word with the sense occurring most frequently in Semcor. In our approach, we will also rely on Semcor as the basic resource, both for training ex- amples and as an indicator of the distribution of the senses of the target word. The goal of our experiment is to evaluate up to which point we can automatically acquire examples for word senses and train accurate supervised WSD systems on them. This is a very promising line of research, but one which remains relatively under- studied (cf. Section 2). The method we applied is based on the monosemous relatives of the target words (Leacock et al., 1998), and we studied some parameters that affect the quality of the acquired corpus, such as the distribution of the number of training instances per each word sense (bias), and the type of features used for disambiguation (local vs. topical). Basically, we built three systems, one fully su- pervised (using examples from both Semcor …

Content truncated to 500 words