UBC-ZAS: A k-NN based Multiclassifier System to perform WSD in a Reduced Dimensional Vector Space

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 358–361, Prague, June 2007. c⃝2007 Association for Computational Linguistics UBC-ZAS: A k-NN based Multiclassifier System to perform WSD in a Reduced Dimensional Vector Space Ana Zelaia, Olatz Arregi and Basilio Sierra Computer Science Faculty University of the Basque Country ana.zelaia@ehu.es Abstract In this article a multiclassifier approach for word sense disambiguation (WSD) prob- lems is presented, where a set of k-NN classifiers is used to predict the category (sense) of each word. In order to combine the predictions generated by the multiclas- sifier, Bayesian voting is applied. Through all the classification process, a reduced di- mensional vector representation obtained by Singular Value Decomposition (SVD) is used. Each word is considered an indepen- dent classification problem, and so differ- ent parameter setting, selected after a tun- ing phase, is applied to each word. The ap- proach has been applied to the lexical sam- ple WSD subtask of SemEval 2007 (task 17). 1 Introduction Word Sense Disambiguation (WSD) is an impor- tant component in many information organization and management tasks. Both, word representation and classification method are crucial steps in the word sense disambiguation process. In this article both issues are considered. On the one hand, Latent Semantic Indexing (LSI) (Deerwester et al., 1990), which is a variant of the vector space model (VSM) (Salton and McGill, 1983), is used in order to ob- tain the vector representation of the corresponding word. This technique compresses vectors represent- ing word related contexts into vectors of a lower- dimensional space. LSI, which is based on Singu- lar Value Decomposition (SVD) (Berry and Browne, 1999) of matrices, has shown to have the ability to extract the relations among features representing words by means of their context of use, and has been successfully applied to Information Retrieval tasks. On the other hand, a multiclassifier (Ho et al., 1994) which uses different training databases is con- structed. These databases are obtained from the original training set by random subsampling. The implementation of this approach is made by a model inspired in bagging (Breiman, 1996), and the k-NN classification algorithm (Dasarathy, 1991) is used to make the sense predictions for testing words. Our group (UBC-ZAS) has participated in the lex- ical sample subtask of SemEval-2007 for task 17, which consists on 100 different words for which a training and testing database have been provided. The aim of this article is to give a brief descrip- tion of our approach to deal with the WSD task and to show the results obtained. In Section 2, our ap- proach is presented. In Section 3, the experimen- tal setup is introduced. The experimental results are presented and discussed in Section 4, and finally, Section 5 contains some conclusions and comments on future work. 2 Proposed Approach In this article a multiclassifier based WSD system which classifies word senses represented in a re- duced dimensional vector space is proposed. In Figure 1 an illustration of the experiment per- formed for each …

Content truncated to 500 words