A comparative study between possibilistic and probabilistic approaches for monolingual word sense disambiguation

This paper proposes and assesses a new possibilistic approach for automatic monolingual word sense disambiguation (WSD). In fact, in spite of their advantages, the traditional dictionaries suffer from the lack of accurate information useful for WSD. Moreover, there exists a lack of high-coverage semantically labeled corpora on which methods of learning could be trained. For these multiple reasons, it became important to use a semantic dictionary of contexts (SDC) ensuring the machine learning in a semantic platform of WSD. Our approach combines traditional dictionaries and labeled corpora to build a SDC and identify the sense of a word by using a possibilistic matching model. Besides, we present and evaluate a second new probabilistic approach for automatic monolingual WSD. This approach uses and extends an existing probabilistic semantic distance to compute similarities between words by exploiting a semantic graph of a traditional dictionary and the SDC. To assess and compare these two approaches, we performed experiments on the standard ROMANSEVAL test collection and we compared our results to some existing French monolingual WSD systems. Experiments showed an encouraging improvement in terms of disambiguation rates of French words. These results reveal the contribution of possibility theory as a mean to treat imprecision in information systems.