GEAM: A General and Event-Related Aspects Model for Twitter Event Detection

摘要

Event detection on Twitter has become a promising research direction due to Twitter's popularity, up-to-date feature, free writing style and so on. Unfortunately, it's a challenge to analyze Twitter dataset for event detection, since the informal expressions of short messages comprise many abbreviations, Internet buzzwords, spelling mistakes, meaningless contents etc. Previous techniques proposed for Twitter event detection mainly focus on clustering bursty words related to the events, while ignoring that these words may not be easily interpreted to clear event descriptions. In this paper, we propose a General and Event-related Aspects Model (GEAM), a new topic model for event detection from Twitter that associates General topics and Event-related Aspects with events. We then introduce a collapsed Gibbs sampling algorithm to estimate the word distributions of General topics and Event-related Aspects in GEAM. Our experiments based on over 7 million tweets demonstrate that GEAM outperforms the state-of-the-art topic model in terms of both Precision and DERate (measuring Duplicated Events Rate detected). Particularly, GEAM can get better event representation by providing a 4-tuple (Time, Locations, Entities, Keywords) structure of the detected events. We show that GEAM not only can be used to effectively detect events but also can be used to analyze event trends.