Abstractive News Summarization based on Event Semantic Link Network

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 236–246, Osaka, Japan, December 11-17 2016. Abstractive News Summarization based on Event Semantic Link Network Wei Li1,+, Lei He2,∗, Hai Zhuge3,+,∗ +Key Lab of Intelligent Information Processing, Institute of Computing Technology, University of Chinese Academy of Sciences, Beijing, China ∗Aston University, Birmingham, UK 1weili.ict.kg@gmail.com, 2hel2@aston.ac.uk, 3zhuge@ict.ac.cn Abstract This paper studies the abstractive multi-document summarization for event-oriented news texts through event information extraction and abstract representation. Fine-grained event mentions and semantic relations between them are extracted to build a unified and connected event seman- tic link network, an abstract representation of source texts. A network reduction algorithm is pro- posed to summarize the most salient and coherent event information. New sentences with good linguistic quality are automatically generated and selected through sentences over-generation and greedy-selection processes. Experimental results on DUC 2006 and DUC 2007 datasets show that our system significantly outperforms the state-of-the-art extractive and abstractive baselines under both pyramid and ROUGE evaluation metrics. 1 Introduction Automatic summarization on news documents enables readers more easily to get general information of interesting news. Most of existing summarization methods have neglected the important event-oriented characteristics of news texts although some popular tasks such as DUC (Document Understanding Con- ference) and TAC (Text Analysis Conference) target at summarizing news documents. The examples below show that the core information of news texts is the atomic event mentions as shown in bolded words and their related concepts as shown in italic phrases. • Lawyer Morris Dees, who is representing Victoria Keenan after she was attacked by two guards in July 1998, introduced depositions to contradict the men’s testimony. • Morris S. Dees Jr., who was the co-founder of the Southern Poverty Law Center, defended for Keenan after she was assaulted by two security guards near the headquarters of the Aryan Nations. An event usually tells us “who did what to whom when and where ...” The most important com- ponents of an event include its actor (who, the agent of the event), action (what, the core meaning of the event) and receiver (whom, the target of the event action). Other arguments indicate other attributes of the event, such as time (when) and location (where). The event arguments are concepts related with the event action. For event-based news summarization, extracting the most salient events and related concepts are the core tasks. One of the most similar related work (Glavaˇs and ˇSnajder, 2014) investigated constructing event graph for multi-document summarization. The nodes in event graph denote event mentions while edges denote temporal relations between event mentions. It ranks event mentions based on the temporal relations and then generates summary by extracting sentences that contain salient event mentions. However, the prob- lems of information overlapping and lacking of coherence cannot be overcomed by extractive methods. This paper explores the issue of abstractive summarization for event-oriented news texts. The semantic relations between events like cause-effect relation are also extracted to help generate more coherent and informative …

Content truncated to 500 words