Learning good decompositions of complex questions

Abstract

This paper proposes a supervised approach for automatically learning good decompositions of complex questions. The training data generation phase mainly builds on three steps to produce a list of simple questions corresponding to a complex question: i) the extraction of the most important sentences from a given set of relevant documents (which contains the answer to the complex question), ii) the simplification of the extracted sentences, and iii) their transformation into questions containing candidate answer terms. Such questions, considered as candidate decompositions, are manually annotated (as good or bad candidates) and used to train a Support Vector Machine (SVM) classifier. Experiments on the DUC data sets prove the effectiveness of our approach. © 2012 Springer-Verlag.