Improving WSD with Multi-Level View of Context Monitored by Similarity Measure

Improving WSD with Multi-Level View of Context Monitored by Similarity Measure E. Crestan<1'2l, M. El-Beze<tl and C. de Loupy<2l (1) Laboratoire d'Infortl1atique d' Avignon 339 ch. Des Meinajaries, BP 1228 F-84911 Avignon Cedex 9 { eric.crestan, marc.elbeze }@lia.univ-avignon.fr Abstract The approach presented in this paper for Word Sense Disambiguation (WSD) is based on a combination of different views of the context. Semantic Classification Trees (SCT) are employed over a short and a multi-level view of context, including rough semantic features, while a similarity measure is used in some particular cases to rely on a larger view of the context. We also describe our two-step approach based on HMM for the all-word task. Introduction In the tracks of SENSEV AL-l (Kilgarriff and Rosenzweig, 2000), the second edition of the word sense disambiguation evaluation campaign offers a new set of words to test improvements in the domain of WSD. It also includes a new task, aimed at disambiguating each word of a text. Our approach for the lexical sample task is based on three different views of the context, which allows us to consider more information for sense tagging. In order to deal with short- range view of the context, we have chosen to use Semantic Classification Trees (SCT) (Kuhn and De Mori, 1995), which are binary decision trees. Moreover, based on our experience, we will show, that using rough semantic features as a higher-level view of the context yields substantial increases in perfom1ance. Finally, a similarity distance is employed in order to capture longer-range context information. The paper is organized as follows: in the first part (Section I), the work we have done on the lexical sample task is presented. This part includes a brief overview of the SCT approach (Section 1. 1) and we show how the coverage it yields could be increased while using more or less rough semantic features thanks to a multi-level view of the context (Section 1.2). In 67 (2) Sinequa 51-59 rue Ledru Rollin F-94200 lvry-sur-Seine { crestan, loupy} @sinequa.com Section 1.3, we propose to use a similarity measure like those used in document retrieval in order to select a sense among those proposed by the SCT systems. The second part (Section 2) is dedicated to the all-words task. A two-step approach based on a trisem-bisem model is presented (Section 2.1 ). Then, we propose to apply a special process on the most frequent words in the task (Section 2.2). In conclusion, the results for both tasks are presented. 1 Lexical Sample Task The lexical sample task of SENSEV AL-2 is composed of 29 nouns, 29 verbs and 15 adjectives in context. We decided to handle the totality of the words, and always assign one and only one sense to each test word (recall= precision). For training purpose, we used the corpus supplied for each word to be disambiguated. However, the number of training sentences supplied was greatly reduced compared to that of the first SENSEV AL exercise. By comparison, the average number of training …

Content truncated to 500 words