Word Sense Disambiguation based on term to term similarity in a context space

Word Sense Disambiguation based on Term to Term Similarity in a Context Space Javier Artiles Dpto. Lenguajes y Sistemas Inform´aticos UNED, Spain javart@bec.uned.es Anselmo Pe˜nas Dpto. Lenguajes y Sistemas Inform´aticos UNED, Spain anselmo@lsi.uned.es Felisa Verdejo Dpto. Lenguajes y Sistemas Inform´aticos UNED, Spain felisa@lsi.uned.es Abstract This paper describes the exemplar based ap- proach presented by UNED at Senseval-3. In- stead of representing contexts as bags of terms and defining a similarity measure between con- texts, we propose to represent terms as bags of contexts and define a similarity measure be- tween terms. Thus, words, lemmas and senses are represented in the same space (the context space), and similarity measures can be defined between them. New contexts are transformed into this representation in order to calculate their similarity to the candidate senses. We show how standard similarity measures obtain better results in this framework. A new similar- ity measure in the context space is proposed for selecting the senses and performing disambigua- tion. Results of this approach at Senseval-3 are here reported. 1 Introduction Word Sense Disambiguation (WSD) is the task of deciding the appropriate sense for a partic- ular use of a polysemous word, given its tex- tual or discursive context. A previous non triv- ial step is to determine the inventory of mean- ings potentially attributable to that word. For this reason, WSD in Senseval is reformulated as a classification problem where a dictionary be- comes the class inventory. The disambiguation process, then, consists in assigning one or more of these classes to the ambiguous word in the given context. The Senseval evaluation forum provides a controlled framework where different WSD systems can be tested and compared. Corpus-based methods have offered encour- aging results in the last years. This kind of methods profits from statistics on a training corpus, and Machine Learning (ML) algorithms to produce a classifier. Learning algorithms can be divided in two main categories: Super- vised (where the correct answer for each piece of training is provided) and Unsupervised (where the training data is given without any answer indication). Tests at Senseval-3 are made in various languages for which two main tasks are proposed: an all-words task and a lexical sam- ple task. Participants have available a training corpus, a set of test examples and a sense inven- tory in each language. The training corpora are available in a labelled and a unlabelled format; the former is mainly for supervised systems and the latter mainly for the unsupervised ones. Several supervised ML algorithms have been applied to WSD (Ide and V´eronis, 1998), (Es- cudero et al., 2000): Decision Lists, Neural Net- works, Bayesian classifiers, Boosting, Exemplar- based learning, etc. We report here the exemplar-based approach developed by UNED and tested at the Senseval-3 competition in the lexical sample tasks for English, Spanish, Cata- lan and Italian. After this brief introduction, Sections 2 and 3 are devoted, respectively, to the training data and the processing performed over these data. Section 4 characterizes the UNED WSD system. First, we …

Content truncated to 500 words