Methods combination and ML-based re-ranking of multiple hypothesis for question-answering systems

Arnaud Grappy (1) , Brigitte Grau (1, 2) , Sophie Rosset (1)

1 LIMSI - Laboratoire d'Informatique pour la Mécanique et les Sciences de l'Ingénieur
2 ENSIIE - Ecole Nationale Supérieure d'Informatique pour l'Industrie et l'Entreprise

Question answering systems answer correctly to different questions because they are based on different strategies. In order to increase the number of questions which can be answered by a single process, we propose solutions to combine two question answering systems, QAVAL and RITEL. QAVAL proceeds by selecting short passages, annotates them by question terms, and then extracts from them answers which are ordered by a machine learning validation process. RITEL develops a multi-level analysis of questions and documents. Answers are extracted and ordered according to two strategies: by exploiting the redundancy of candidates and a Bayesian model. In order to merge the system results, we developed different methods either by merging passages before answer ordering, or by merging end-results. The fusion of end-results is realized by voting, merging, and by a machine learning process on answer characteristics, which lead to an improvement of the best system results of 19 %.

Domaines: Informatique [cs]; Informatique et langage [cs.CL]

Arnaud Grappy, Brigitte Grau, Sophie Rosset. Methods combination and ML-based re-ranking of multiple hypothesis for question-answering systems. Proceedings of the Workshop on Innovative Hybrid Approaches to the Processing of Textual Data, EACL, Apr 2012, Avignon, France. pp.87--96. ⟨hal-02289723⟩