Subtopic-driven Multi-Document Summarization

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 3153–3162, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 3153 Subtopic-driven Multi-Document Summarization Xin Zheng1,2, Aixin Sun1, Jing Li3, Karthik Muthuswamy2 1 School of Computer Science and Engineering, Nanyang Technological University, Singapore 2 SAP Innovation Center Network, SAP Asia Pte Ltd, Singapore 3 Inception Institute of Artificial Intelligence Abu Dhabi, United Arab Emirates xzheng008@e.ntu.edu.sg, axsun@ntu.edu.sg, jingli.phd@hotmail.com, karthik.muthuswamy@sap.com Abstract In multi-document summarization, a set of documents to be summarized is assumed to be on the same topic, known as the under- lying topic in this paper. That is, the un- derlying topic can be collectively represented by all the documents in the set. Meanwhile, different documents may cover various differ- ent subtopics and the same subtopic can be across several documents. Inspired by topic model, the underlying topic of a document set can also be viewed as a collection of dif- ferent subtopics of different importance. In this paper, we propose a summarization model called STDS. The model generates the under- lying topic representation from both document view and subtopic view in parallel. The learn- ing objective is to minimize the distance be- tween the representations learned from the two views. The contextual information is encoded through a hierarchical RNN architecture. Sen- tence salience is estimated in a hierarchical way with subtopic salience and relative sen- tence salience, by considering the contextual information. Top ranked sentences are then extracted as a summary. Note that the notion of subtopic enables us to bring in additional information (e.g., comments to news articles) that is helpful for document summarization. Experimental results show that the proposed solution outperforms state-of-the-art methods on benchmark datasets. 1 Introduction Multi-document summarization (MDS) is useful in many applications, e.g., summarizing answers in forums (Song et al., 2017) and reports of burst events (Kedzie et al., 2016). In MDS, the doc- uments in a set are assumed to share the same underlying topic. Given a set of documents to be summarized, the important information is col- lectively determined by all the documents in the set, rather than simple integration of key points in each document. Hence, the correlations among documents become crucial for identifying the im- portant information to be included in a summary. Moreover, each sentence should not be interpreted independently. Its context (or the surrounding sentences) does affect the information expressed in a sentence (Nenkova et al., 2006; Ren et al., 2017). Therefore, both correlations among doc- uments and contextual information within a single document should be considered in MDS. Various methods have been developed for MDS. Graph-based models (Mihalcea and Tarau, 2004; Erkan and Radev, 2004) blend sentences from dif- ferent documents together and attempt to lever- age the correlations among documents to extract the most representative sentences. Nonetheless these methods are less effective when the sentence graph is not well connected. Thus, many stud- ies attempt to construct well connected graphs, …

Content truncated to 500 words