A Distant Learning Approach for Extracting Hypernym Relations from Wikipedia Disambiguation Pages

Extracting hypernym relations from text is one of the key steps in the automated construction and enrichment of semantic resources. The state of the art offers a large varierty of methods (linguistic, statistical, learning based, hybrid). This variety could be an answer to the need to process each corpus or text fragment according to its specificities (e.g. domain granularity, nature, language, or target semantic resource). Moreover, hypernym relation may take different linguistic forms. The aim of this paper is to study the behaviour of a supervised learning approach to extract hypernym relations whatever the way they are expressed, and to evaluate its ability to capture regularities from the corpus, without human intervention. We apply a distant supervised learning algorithm on a sub-set of Wikipedia in French made of disambiguation pages where we manually annotated hypernym relations. The learned model obtained a F-measure of 0.67, outperforming lexico-syntactic pattern matching used as baseline.