This paper proposes an abstractive multi-document summarization method. Given a document set, the system first generates sentence clusters through an event clustering algorithm using distributed representation. Each cluster is regarded as a subtopic of this set. Then we use a novel multi-sentence compression method to generate K-shortest paths for each cluster. Finally, some preferable paths are selected from these candidates to construct the final summary based on several customized submodular functions, which are designed to measure the summary quality from different perspectives. Experimental results on DUC 2005 and DUC 2007 datasets demonstrate that our method achieves better performance compared with the state-of-the-art systems.
展开▼
