News Stream Summarization using Burst Information Networks

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 784–794, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics News Stream Summarization using Burst Information Networks Tao Ge1,2∗, Lei Cui3, Baobao Chang1,2, Sujian Li1,2, Ming Zhou3, Zhifang Sui1,2 1Key Laboratory of Computational Linguistics, Ministry of Education, School of EECS, Peking University, Beijing, 100871, China 2Collaborative Innovation Center for Language Ability, Xuzhou, Jiangsu, 221009, China 3Microsoft Research getao@pku.edu.cn, lecu@microsoft.com, chbb@pku.edu.cn lisujian@pku.edu.cn, mingzhou@microsoft.com, szf@pku.edu.cn Abstract This paper studies summarizing key informa- tion from news streams. We propose sim- ple yet effective models to solve the problem based on a novel and promising representation of text streams – Burst Information Networks (BINets). A BINet can be aware of redundant information, allows global analysis of a text stream, and can be efficiently built and dy- namically updated, which perfectly fits the de- mands of text stream summarization. Exten- sive experiments show that the BINet-based approaches are not only efficient and can be used in a real-time online summarization set- ting, but also can generate high-quality sum- maries, outperforming the state-of-the-art ap- proach. 1 Introduction Text stream summarization aims to summarize key information from a text stream containing huge numbers of documents, which is an important and useful task that can be used for many real-world ap- plications. For example, a news portal website ed- itor needs to summarize news streams in the past day for generating a list of headline news; an edi- tor of Sports Weekly may want a summary of the past week news stream for editing the magazine; and geologists and meteorologists will benefit from a summary of disaster events from the past year news stream (as shown in Table 1) for their study. In contrast to traditional text summarization tasks (e.g., single and multi-document summarization) ∗ This work was done when the first author was visiting Microsoft Research Asia single-document summarization multi-document summarization stream summarization = sentence = document Figure 1: Stream summarization paradigm. that have been extensively studied for decades, the task of stream summarization is a younger research problem which attempts to solve a summarization problem in the big-data setting. For a text stream with millions of documents involving various topics and events, traditional single- and multi-document summarization approaches cannot address the infor- mation overload challenge. For example, a single- document summarization model will generate 1 mil- lion document summaries for a text stream with 1 million documents, which are still overwhelming for a person to learn the key information in the stream. In such cases, one needs to a summary of the whole stream instead of summaries of each document. Figure 1 shows the paradigm of stream sum- marization. Compared with single- and multi- document summarization, stream summarization has three differences: (1) it summarizes a text stream containing millions of documents involving a variety of topics and events while single- and 784 2009 disaster summary 2010 disaster summary • ... • ... • Sep 2, 2009: About 60 people die when a …

Content truncated to 500 words