The R2D2 team at Senseval-3

The R2D2 Team at SENSEVAL-3∗ Sonia V´azquez, Rafael Romero Armando Su´arez and Andr´es Montoyo Dpto. de Lenguajes y Sistemas. Inform´aticos Universidad de Alicante, Spain {svazquez,romero}@dlsi.ua.es {armando,montoyo}@dlsi.ua.es Manuel Garc´ıa, M. Teresa Mart´ın † M. ´Angel Garc´ıa and L. Alfonso Ure˜na Dpto. de Inform´atica Universidad de Ja´en, Spain {mgarcia,maite}@ujaen.es {magc,laurena}@ujaen.es Davide Buscaldi, Paolo Rosso ‡ Antonio Molina, Ferr´an Pl´a and Encarna Segarra Dpto. de Sistemas Inform´aticos y Computaci´on Univ. Polit. de Valencia, Spain {dbuscaldi,prosso}@dsic.upv.es {amolina,fpla,esegarra}@dsic.upv.es Abstract The R2D2 systems for the English All-Words and Lexical Sample tasks at SENSEVAL-3 are based on several supervised and unsupervised methods com- bined by means of a voting procedure. Main goal was to take advantage of training data when avail- able, and getting maximum coverage with the help of methods that not need such learning examples. The results reported in this paper show that super- vised and unsupervised methods working in par- allel, and a simple sequence of preferences when comparing the answers of such methods, is a feasi- ble method... The whole system is, in fact, a cascade of deci- sions of what label to assign to a concrete instance based on the agreement of pairs of systems, when it is possible, or selecting the available answer from one of them. In this way, supervised are preferred to unsupervised methods, but these last ones are able to tag such words that not have available training data. 1 Introduction Designing a system for Natural Language Process- ing (NLP) requires a large knowledge on language structure, morphology, syntax, semantics and prag- matic nuances. All of these different linguistic knowledge forms, however, have a common asso- ciated problem, their many ambiguities, which are difficult to resolve. In this paper we concentrate on the resolution of the lexical ambiguity that appears when a given word in a context has several different meanings. ∗ This paper has been partially supported by the Spanish Government (CICyT) under project number TIC-2003-7180 and the Valencia Government (OCyT) under project number CTIDIB-2002-151 This specific task is commonly referred as Word Sense Disambiguation (WSD). This is a difficult problem that is receiving a great deal of attention from the research community because its resolu- tion can help other NLP applications as Machine Translation (MT), Information Retrieval (IR), Text Processing, Grammatical Analysis, Information Ex- traction (IE), hypertext navigation and so on. The R2D2 Team has participated in two tasks: English all-words and lexical sample. We use sev- eral different systems both supervised and unsuper- vised. The supervised methods are based on Max- imum Entropy (ME) (Lau et al., 1993; Berger et al., 1996; Ratnaparkhi, 1998), neural network using the Learning Vector Quantization algorithm (Koho- nen, 1995) and Specialized Hidden Markov Mod- els (Pla, 2000). The unsupervised methods are Rel- evant Domains (RD) (Montoyo et al., 2003) and the CIAOSENSO WSD system which is based on Conceptual Density (Agirre and Rigau, 1995), fre- quency of WordNet (Miller et al., 1993a) senses and WordNet Domains (Magnini and Cavaglia, 2000). In the following section we will show a more …

Content truncated to 500 words