Section: New Results

Extending wordnets

Participants : Benoît Sagot, Marianna Apidianaki, Valérie Hanoka.

The WOLF (see section  5.9 ) is a freely available, automatically created wordnet for French, the biggest drawback of which has until now been the lack of general concepts that are typically expressed with highly polysemous vocabulary that is on the one hand the most valuable for applications in human language technologies but also the most difficult to add to wordnet accurately with automatic methods on the other. In collaboration with Darja Fišer (University of Ljubljana), we have developed a self-training-like technique for acquiring a classifier that is able to assign appropriate synset ids (i.e., senses) to new words, extracted from non-disambituated multilingual sources of lexical knowledge, such as Wiktionaries and Wikipedia [39] , [40] . Automatic and manual evaluation shows high coverage as well as high quality of the resulting lexico-semantic repository. Another important advantage of the approach is that it is fully automatic and language-independent and can therefore be applied to any other language still lacking a wordnet. Indeed, it was applied to Slovene as well.

Other techniques were used as well and are the basis of various submitted conference papers. They rely, among others, on morphological derivation, on graph-based representation of highly multilingual lexicons extracted from numerous wiktionaries, and on automatically induced sense clusters.