Répondre à des questions à réponses multiples : premières expérimentations

Question-answering systems have multiples problems dealing with answer extraction from retrieved documents, especially when cross-checking is needed for questions expecting multiple answers. We focused on answer extraction and answer crossing by building an ideal corpus for a dozen of multiple answer questions. We performed a fine-grained evaluation on each strategy for answer extraction and validation. We compare the results of two questionanswering systems : FIDJI which has successfully participated to several evaluation campaigns and Citron which is designed for multiple answer extraction and cross-checking. MOTS-CLÉS : question-réponse, questions à réponses multiples, question liste