TopicDSDR: combining topic decomposition and data reconstruction for summarization

Zhiming Zhang,Hongjie Li,Lian'en Huang

DOI: https://doi.org/10.1007/978-3-642-38562-9_35

2013-01-01

Abstract:Multi-document summarization attempts to select the most important information to generate a compressed summary from a collection of documents. From the perspective of data reconstruction, a good summary may also well reconstruct the original documents. A document generally contains a variety of information centered around a main topic and covers different aspects of the main topic. In this paper we propose a novel model that combines data reconstruction and topic decomposition to summarize the documents, named TopicDSDR, which can not only best reconstruct the original documents but also capture the semantic similarity and main topics. We discuss two kinds of reconstructions: linear reconstruction and nonnegative reconstruction. We use the generalized Kullback-Leibler(KL) divergence as the loss function to evaluate the quality of summary for linear and nonnegative reconstruction and develop two new algorithms respectively. We conduct experiments on DUC2006 and DUC2007 summarization data sets, the experimental results demonstrate the effectiveness of our proposed methods.