The UNED Systems at SENSEVAL-2

The UNED systems at SENSEVAL-2 David Fermindez-Amor6s, Julio Gonzalo, Felisa Verdejo Depto. de Lenguajes y Sistemas Informaticos, UNED { david,julio,felisa }@lsi.uned.es . Abstract We have participated in the SENSEVAL-2 En- glish tasks (all words and lexical sample) with an unsupervised system based on mutual infor- mation measured over a large corpus (277 mil- lion words) and some additional heuristics. A supervised extension of the system was also pre- sented to the lexical sample task. Our system scored first among unsupervised systems in both tasks: 56.9% recall in all words, 40.2% in lexical sample. This is slightly worse than the first sense heuristic for all. words and 3.6% better for the lexical sample, a strong in- dication that unsupervised Word Sense Disam- biguation remains being a strong challenge. 1 Introduction We advocate researching unsupervised tech- niques for Word Sense Disambiguation (WSD). Supervised techniques offer better results in general but the setbacks, such as the problem of developing reliable training data, are very considerable. Also there's probably more to WSD than blind machine learning (a typical ap- proach, although such systems produce interest- ing baselines). Within the unsupervised paradigm, we are in- terested in performing in-depth measures of the disambiguation potential of different sources of information. We have previously investigated the informational value of semantic distance measures in (Fermindez-Amor6s et al., ) . For SENSEVAL-2, we have turned to investigate pure coocurrence information as a source of disam- biguation evidence. In essence, our system com- putes a matrix of mutual information for a fixed vocabulary and applies it to weight coocurrence counting between sense and context character- istic vectors. 75 In the next section we describe the process of constructing the relevance matrix. In section 3 we present the particular heuristics used for the competing systems. In section 4 we show the results by system and heuristic and some base- lines for comparison. Finally in the last sections we draw some conclusions about the exercise. 2 2.1 The Relevance matrix Corpus processing Before building our systems we have developed a resource we've called the relevance matrix. The raw data used to build the matrix comes from the Project Gutenberg (PG) 1. At the time of the creation of the matrix the PG consisted of more than 3000 books of di- verse genres. We have adapted these books for our purpose : First, language identification was used to filter books written in English; Then we stripped off the disclaimers. The result is a collection of around 1.3Gb of plain text. Finally we tokenize, lemmatize, strip punctu- ation and stop words and detect numbers and proper nouns. 2.2 Coocurrence matrix We have built a vocabulary of the 20000 most frequent words (or labels, as we have changed all the proper nouns detected to the label PROPER_NOUN and all numbers detected to NUMBER) in the text and a symmetric coocur- rence matrix between these words within a con- text of 61 words (we thought a broad context of radius 30 would be appropriate since we …

Content truncated to 500 words