The University of Jaén Word Sense Disambiguation system

The University of Jaén Word Sense Disambiguation System* Manuel García-Vega Universidad de Jaén Av. Madrid 35 Jaén, Spain, 23071 mgarcia@ujaen.es Miguel A. García-Cumbreras Universidad de Jaén Jaén, Spain, 23071 Av. Madrid 35, 23071 magc@ujaen.es M. Teresa Martín-Valdivia Universidad de Jaén Av. Madrid 35 Jaén, Spain, 23071 maite@ujaen.es L. Alfonso Ureña-López Universidad de Jaén Av. Madrid 35 Jaén, Spain, 23071 laurena@ujaen.es * This paper has been partially supported by the Spanish Government (MCYT) Project number TIC2003-07158- C04-04 Abstract This paper describes the architecture and re- sults of the University of Jaén system pre- sented at the SENSEVAL-3 for the English- lexical-sample and English-All-Words tasks. The system is based on a neural network ap- proach. We have used the Learning Vector Quantization, which is a supervised learning algorithm based on the Kohonen neural model. 1 Introduction Our system for SENSEVAL-3 uses a supervised learning algorithm for word sense disambiguation. The method suggested trains a neural network us- ing the Learning Vector Quantization (LVQ) algo- rithm, integrating several semantic relations of WordNet (Fellbaum, 1998) and SemCor corpus (Miller et al., 1993). The University of Jaén system has been used in English-lexical-sample and Eng- lish-All-Words tasks. 2 Experimental Environment The presented disambiguator uses the Vector Space Model (VSM) as an information representa- tion model. Each sense of a word is represented as a vector in an n-dimensional space where n is the number of words in all its contexts. The accuracy of the disambiguator depends es- sentially on the word weights. We use the LVQ algorithm to adjust them. The input vector weights are calculated as shown by (Salton and McGill, 1983) with the standard tf·idf, where the documents are the paragraphs. They are presented to the LVQ network and, after training, the output vectors (called prototype or codebook vectors) are ob- tained, containing the adjusted weights for all senses of each word. Any word to disambiguate is represented with a vector in the same way. This representation must be compared with all the trained word sense vec- tors by applying the cosine similarity rule: i k i k i k sim x w wx x w ·· ) , ( = [1] The sense corresponding to the vector of highest similarity is selected as the disambiguated sense. To train the neural network we have integrated semantic information from two linguistic re- sources: SemCor corpus and WordNet lexical da- tabase. 2.1 SemCor Firstly, the SemCor (the Brown Corpus labeled with the WordNet senses) was fully used (the Brown-1, Brown-2 and Brown-v partitions). We used the paragraph as a contextual semantic unit Association for Computational Linguistics for the Semantic Analysis of Text, Barcelona, Spain, July 2004 SENSEVAL-3: Third International Workshop on the Evaluation of Systems and each context was included in the training vec- tor set. The SENSEVAL-3 English tasks have used the WordNet 1.7.1 sense inventory, but the SemCor is tagged with an earlier version of WordNet (spe- cifically WordNet version 1.6). Figure 1. SemCor context for “climb”. Therefore it was necessary to update …

Content truncated to 500 words