Rol semantikoen etiketatze automatikoarol multzoak eta hautapen murriztapenak

Résumé

This thesis focuses on two well-known open issues in Semantic Role Classi fication (SRC) research: (1) the suitability of diferent role inventories in practice, and (2) the limited in uence and sparseness of lexical features. About the former, we present an empirical comparative study on the use of PropBank vs. VerbNet roles, the two most widely used role inventories, testing the performance diferences for unseen verbs and the robustness for new corpus domains. About the latter, we test the use of automatically learnt selectional preferences as a complement to lexical features, proposing both WordNet-based and distributional similarity based models. We show that all our selectional preference models improve over lexical features in in-vitro experiments, and that the models are complementary. Finally, we show that incorporating features based on selectional preferences, the overall performance of an state-of-the-art SRC system improves both in in-domain and out-of-domain corpora.