Support vector machines for query-focused summarization trained and evaluated on pyramid data

Abstract

This paper presents the use of Support Vector Machines (SVM) to detect relevant information to be included in a queryfocused summary. Several classifiers are trained using pyramids of summary content units information. The Mapping-Convergence algorithm is used with positive, unlabeled data, and a small set of negative seeds. The SVMs are tested on two Document Understanding Conference (DUC) 2006 systems. The performance of the new approaches is compared with the original systems using the DUC 2005 corpus as test data. For evaluation purposes, we also present an automatic method based on pyramid data with good correlation with other human or automatic procedures.