Graph based Multi-Document Summarization with Latent Topics

Abstract

The necessity of automatic text summarization has been recently increasing for helping people choose their necessary information. While various methods have been proposed, the usefulness of graph-based text summarization methods such as LexRank is recognized. LexRank computes the importance of sentences based on the idea of eigenvector centrality in a graph representation of sentences. This method uses the surface information to measure similarity among sentences such as cosine similarity, but does not use latent topic similarity. In this study, we propose a multi-document summarization method using a graph of sentences based on latent topics and show our method can summarize multiple documents with higher accuracy than LexRank through an experiment using DUC2004 task.