Generic Multi-Document Summarization Using Topic-Oriented Information

The graph-based ranking models have been widely used for multi-document summarization recently. By utilizing the correlations between sentences, the salient sentences can be extracted according to the ranking scores. However, sentences are treated in a uniform way without considering the topic-level information in traditional methods. This paper proposes the topic-oriented PageRank (ToPageRank) model, in which topic information is fully incorporated, and the topic-oriented HITS (ToHITS) model is designed to compare the influence of different graph-based algorithms. We choose the DUC2004 data set to examine the models. Experimental results demonstrate the effectiveness of ToPageRank. And the results also show that ToPageRank is more effective and robust than other models including ToHIST under different evaluation metrics.