TKB-UO: Using Sense Clustering for WSD

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 322–325, Prague, June 2007. c⃝2007 Association for Computational Linguistics TKB-UO: Using Sense Clustering for WSD Henry Anaya-S´anchez1, Aurora Pons-Porrata1, Rafael Berlanga-Llavori2 1 Center of Pattern Recognition and Data Mining, Universidad de Oriente, Cuba 2 Computer Science, Universitat Jaume I, Spain 1 {henry,aurora}@csd.uo.edu.cu 2 berlanga@lsi.uji.es Abstract This paper describes the clustering-based approach to Word Sense Disambiguation that is followed by the TKB-UO system at SemEval-2007. The underlying disambigua- tion method only uses WordNet as external resource, and does not use training data. Re- sults obtained in both Coarse-grained En- glish all-words task (task 7) and English fine-grained all-words subtask (task 17) are presented. 1 Introduction The TKB-UO system relies on the knowledge- driven approach to Word Sense Disambiguation (WSD) presented in (Anaya-S´anchez et al., 2006). Regarding that meaningful senses of words in a tex- tual unit must be coherently related, our proposal uses sense clustering with the aim of determining cohesive groups of senses that reflect the connectiv- ity of the disambiguating words. The way this proposal uses clustering for disam- biguation purposes is different from those usages re- ported in other works of the WSD area. For ex- ample, in (Pedersen et al., 2005) textual contexts are clustered in order to represent senses for Word Sense Discrimination. Other works like (Agirre and L´opez, 2003), cluster fine-grained word senses into coarse-grained ones for polysemy reduction. In- stead, our method clusters all possible senses cor- responding to all words in a disambiguating textual unit. Thus, our system implements a novel cluster- ing approach for the contextual disambiguation of words. We use the lexical resource WordNet (version 2.1) as the repository of word senses, and also as the provider of sense representations. It is worth men- tioning that our proposal does not require the use of training data. 2 The disambiguation algorithm Our method starts with a clustering of all possible senses of the disambiguating words. Such a cluster- ing tries to identify cohesive groups of word senses, which are assumed to represent the different mean- ings for the set of disambiguating words. Then, clus- ters that match the best with the context are selected via a filtering process. If the selected clusters dis- ambiguate all words, the process is stopped and the senses belonging to the selected clusters are inter- preted as the disambiguating ones. Otherwise, the clustering and filtering steps are performed again (regarding the remaining senses) until the disam- biguation is achieved. Algorithm 1 shows the general steps of our pro- posal for the disambiguation of a set of words W. In the algorithm, clustering represents the basic clus- tering method, filter is the function that selects the clusters, and T denotes the intended textual context from which words in W are disambiguated (typi- cally, a broader bag of words than W). Next subsec- tions describe in detail each component of the whole process. 2.1 Sense Representation For clustering purposes, word senses are repre- sented as topic signatures (Lin …

Content truncated to 500 words