Feature-Rich Segment-Based News Event Detection on Twitter

International Joint Conference on Natural Language Processing, pages 302–310, Nagoya, Japan, 14-18 October 2013. Feature-Rich Segment-Based News Event Detection on Twitter Yanxia Qin1 Yue Zhang2 Min Zhang3,1 Dequan Zheng1 1 School of Computer Science and Technology, Harbin Institute of Technology, Harbin 150001, China 2 Singapore University of Technology and Design, Singapore 138682 3 School of Computer Science and Technology, SooChow University, Suzhou 215006, China {yxqin,dqzheng}@mtlab.hit.edu.cn yue_zhang@sutd.edu.sg zhangminmt@hotmail.com Abstract Event detection on Twitter is an important and challenging research topic. On the one hand, Twitter provides first-hand informa- tion and fast broadcasting. On the oth- er, challenges include short and noisy con- tent, big volume data and fast-changing topics. Dominant approaches for Twit- ter event detection model events by clus- tering tweets, words or segments, while segments have been proven to be advanta- geous over both words and tweets in news event detection. We study segment-based news event detection, for which existing heuristic-based methods suffer from low recall. We propose feature-based event fil- tering to address this issue. Our filter in- corporate a rich family of features that are empirically proven to be valuable. Exper- imental results show that our event detec- tion system outperforms the state-of-the- art baseline with doubled recall and in- creased precision. 1 Introduction We study news event detection from Twitter mes- sages (tweets). Generally, tweets can be classi- fied into three groups: 1) news events, or breaking news such as “Manchester united Vs Athletic in Jan. 1st”; 2) hot topics that spread among a large amount of Twitter users, such as horoscope top- ics (e.g. “You have recently experienced a phase of expansionism and it’s... More for Sagittarius”); and 3) heterogeneous collections or, meaningless non-event tweets, such as “Need buddy wanna chat”. Some previous work (Cataldi et al., 2010; Kasiviswanathan et al., 2011; Diao et al., 2012) regards both news events and hot topics as sub- jects of detection, while other work (Jackoway et al., 2011; Sakaki et al., 2010; Becker et al., 2012) only detects news events. We are interested in the latter, for which most previous work detects only specific types of events. For example, Sakaki et al. (2010) detect earthquake events from Twitter. In this paper, we study event detection in general. Compared with event detection in news texts, Twitter provides more opportunities and chal- lenges. Yom-Tov and Diaz (2011) report that Twit- ter can broadcast news faster than traditional me- dia, which provides an opportunity for event de- tection in Twitter. On the other hand, there are challenges in event detection from Twitter data: 1) tweets are too short and sometimes cannot carry enough information; 2) tweets contain many noisy words, which can be harmful for event detection and 3) the volume of Twitter data is very large, which makes event detection a big data problem. The dominant approach for Twitter event detec- tion is clustering. Similar tweets (Becker et al., 2011; Li et al., 2012c) or words (Platakis et al., 2009; Lee et al., 2012) are group into a cluster, be- …

Content truncated to 500 words