Event Detection with Burst Information Networks

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 3276–3286, Osaka, Japan, December 11-17 2016. Event Detection with Burst Information Networks Tao Ge1,2∗, Lei Cui3, Baobao Chang1,2, Zhifang Sui1,2, Ming Zhou3 1Key Laboratory of Computational Linguistics, Ministry of Education, School of Electronics Engineering and Computer Science, Peking University, Beijing, China 2Collaborative Innovation Center for Language Ability, Xuzhou, Jiangsu, China 3Microsoft Research, Beijing, China getao@pku.edu.cn, lecu@microsoft.com, chbb@pku.edu.cn szf@pku.edu.cn, mingzhou@microsoft.com Abstract Retrospective event detection is an important task for discovering previously unidentified events in a text stream. In this paper, we propose two fast centroid-aware event detection models based on a novel text stream representation – Burst Information Networks (BINets) for addressing the challenge, following the D2N2K (Data-to-Network-to-Knowledge) paradigm. The BINets are time-aware, efficient and can be easily analyzed for identifying key information (centroids). These advantages allow the BINet-based approaches to achieve the state-of-the-art performance on multiple datasets, demonstrating the efficacy of BINets for the task of event detection. 1 Introduction Retrospective Event Detection (RED) (sometimes called topic detection) is a core task for text stream analysis, which aims to detect events that are previously unknown to the system (Wayne, 1998; Ra- jaraman and Tan, 2001) and is useful for many applications such as text stream summarization and evolutionary analysis of events in both news and social streams. docid time text d1 Jan 12, 2010 A 7.0 magnitude quake hits the impoverished Caribbean nation of Haiti, killing more than 200,000 people, injuring an estimated 300,000. d2 Feb 27, 2010 A huge magnitude 8.8 earthquake strikes near the coast of south-central Chile, shaking buildings, causing blackouts and killing at least 147 people. d3 Apr 14, 2010 A 7.1-magnitude earthquake struck Tibetan Autonomous Prefecture of Yushu in southern Qinghai Province on April 14, 2010, killing at least 400 people and injuring more than 10,000. Table 1: Documents discussing different earthquake events. Most previous event detection approaches tend to use document- or keyword-based clustering models. Another solution proposed in recent years is to build a keyword graph to model the co-occurrence of keywords for detecting keyword communities as events (Sayyadi and Raschid, 2013). Even though these methods can achieve fair performance in small datasets, they have either of the following limitations: • No time-awareness: many event detection models do not take into account time information. As a result, it is very likely that the documents that talk about different events (as Table 1 shows) are grouped into one cluster just because their lexical similarity is high. • Inefficient: clustering-based methods tend to be time-consuming. For example, the time complexity of GAC (group average clustering) – the most commonly used clustering method in event detection – is O(n2 log n). The computational challenge makes them difficult to work on a large dataset. ∗This work was done when the first author was visiting Microsoft Research Asia This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http:// creativecommons.org/licenses/by/4.0/ 3276 • Deviation of cluster centroids: it is likely that …

Content truncated to 500 words