A Participant-based Approach for Event Summarization Using Twitter Streams

Proceedings of NAACL-HLT 2013, pages 1152–1162, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics A Participant-based Approach for Event Summarization Using Twitter Streams Chao Shen1, Fei Liu2, Fuliang Weng2, Tao Li1 1School of Computing and Information Sciences, Florida International University Miami, Florida 33199, USA 2Research and Technology Center, Robert Bosch LLC Palo Alto, California 94304, USA {cshen001, taoli}@cs.fiu.edu {fei.liu, fuliang.weng}@us.bosch.com Abstract Twitter offers an unprecedented advantage on live reporting of the events happening around the world. However, summarizing the Twit- ter event has been a challenging task that was not fully explored in the past. In this paper, we propose a participant-based event summa- rization approach that “zooms-in” the Twit- ter event streams to the participant level, de- tects the important sub-events associated with each participant using a novel mixture model that combines the “burstiness” and “cohesive- ness” properties of the event tweets, and gen- erates the event summaries progressively. We evaluate the proposed approach on different event types. Results show that the participant- based approach can effectively capture the sub-events that have otherwise been shadowed by the long-tail of other dominant sub-events, yielding summaries with considerably better coverage than the state-of-the-art. 1 Introduction Twitter has increasingly become a critical source of information. People report the events they are ex- periencing or publish comments on a wide variety of events happening around the world, ranging from the unexpected natural disasters, regional riots, to many scheduled events, such as sports games, po- litical debates, local festivals, and even academic conferences. The Twitter data streams thus cover a broad range of events and broadcast these in- formation in a live manner. Event summarization in this paper aims to generate a representative and concise textual description of the scheduled events that are being lively reported on Twitter, providing people with an alternative means of observing the world beyond the traditional journalism. Specifi- cally, we investigate scheduled events of different types, including six of the NBA (National Basket- ball Association) sports games and a representative conference event, namely the Apple CEO’s keynote speech in the Apple Worldwide Developers Confer- ence (WWDC 2012)1. All these events have excited great discussion among the Twitter community. Summarizing the Twitter event is a challenging task that has yet been fully explored in the past. Most previous summarization studies focus on the well-formatted news documents, as driven by the annual DUC2 and TAC3 evaluations. In contrast, the Twitter messages (a.k.a., tweets) are very short and noisy, containing nonstandard terms such as ab- breviations, acronyms, emoticons, etc. (Liu et al., 2011b; Liu et al., 2012; Eisenstein, 2013). The noisy contents also cause great difficulties to the tra- ditional NLP tools such as NER and dependency parser (Ritter et al., 2011; Foster et al., 2011), lim- iting the possibility of applying finer-grained event analysis tools. In nature, the event tweets are closely associated with the timeline and are drastically dif- ferent from a static collection of news documents. The tweets converge into text streams that pulse along the timeline …

Content truncated to 500 words