Affinity-Preserving Random Walk for Multi-Document Summarization

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 210–220 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Affinity-Preserving Random Walk for Multi-Document Summarization Kexiang Wang, Tianyu Liu, Zhifang Sui and Baobao Chang Key Laboratory of Computational Linguistics, Ministry of Education School of Electronics Engineering and Computer Science, Peking University Collaborative Innovation Center for Language Ability, Xuzhou 221009 China {wkx,tianyu0421,szf,chbb}@pku.edu.cn Abstract Multi-document summarization provides users with a short text that summarizes the information in a set of related doc- uments. This paper introduces affinity- preserving random walk to the summa- rization task, which preserves the affin- ity relations of sentences by an absorb- ing random walk model. Meanwhile, we put forward adjustable affinity-preserving random walk to enforce the diversity con- straint of summarization in the random walk process. The ROUGE evaluations on DUC 2003 topic-focused summarization task and DUC 2004 generic summariza- tion task show the good performance of our method, which has the best ROUGE- 2 recall among the graph-based ranking methods. 1 Introduction Multi-document summarization provides users with summary that reflects the main information in a set of given documents. The documents are often related and talk about more than one top- ics. Generic multi-document summarization and topic-focused multi-document summarization are two typical kinds of summarization. The former is a summarization delivering the main information of the documents with no bias while the latter is a one delivering the main information biased to a given topic description (a few sentences or even phrases). Most existing summarization systems are designed for these two kinds of summariza- tion. There are two goals for generic multi-document summarization. The first one is saliency. Sum- mary sentences should be central sentences that capture the majority of information in a docu- ment cluster. The sentences with little informa- tion about the document cluster should not be in- cluded in the summary. The second one is di- versity. The information overlap between sum- mary sentences should be as minimal as possi- ble due to the length limit of summary. In other words, the information coverage of summary is a determinant, which requires that the summary sentences should cover diverse aspects of infor- mation. Besides the two goals, there is another goal for the topic-focused summarization and that is relevancy. It requires that the summary sen- tences be relevant to the topic description. A se- ries of conferences and workshops on automatic text summarization (e.g. NTCIR, DUC), special topic sessions in ACL, EMNLP and SIGIR have advanced the techniques to achieve these goals and many approaches have been proposed so far. In this paper, we focus on the extractive summa- rization methods, which extract the summary sen- tences from the input document cluster. We pro- pose affinity-preserving random walk for multi- document summarization. The method is a graph- based ranking method, which takes into account the global information collectively computed from the entire sentence affinity graph. Different from the previous graph-based ranking methods, our method adopts “global normalization” to trans- …

Content truncated to 500 words