Extracting hypernym relations from Wikipedia disambiguation pages : comparing symbolic and machine learning approaches

Extracting hypernym relations from text is one of the key steps in the construction and enrichment of semantic resources. Several methods have been exploited in a variety of propositions in the literature. However, the strengths of each approach on a same corpus are still poorly identified in order to better take advantage of their complementarity. In this paper, we study how complementary two approaches of different nature are when identifying hypernym relations on a structured corpus containing both well-written text and syntactically poor formulations, together with a rich formatting. A symbolic approach based on lexico-syntactic patterns and a statistical approach using a supervised learning method are applied to a sub-corpus of Wikipedia in French, composed of disambiguation pages. These pages, particularly rich in hypernym relations, contain both kinks of formulations. We compared the results of each approach independently of each other and compared the performance when combining together their individual results. We obtain the best results in the latter case, with an F-measure of 0.75. In addition, 55% of the relations identified by our approach, with respect to a reference corpus, are not expressed in the French DBPedia and could be used to enrich this resource.