Twitter Event Detection Based On Topic Models

With the rapid growth of Twitter in recent years, millions of people usethis social network web service to gather real-time news, update personalstatuses or share opinions. Event detection on Twitter has become a promisingresearch topic and attracted more and more attention. Twitter’s popularity, up-to-date feature, free writing style properties make it a better source for eventdetection. Unfortunately, it also brings several challenges such as the informalexpressions, spelling mistakes, Internet buzzwords, abbreviations,meaningless contents etc.In this paper, we propose a General and Event-related Aspects Model(GEAM) for event detection from Twitter. It’s a new topic model thatassociates General topics and Event-related Aspects with events. We thenintroduce a collapsed Gibbs sampling algorithm to estimate the worddistributions of General topics and Event-related Aspects in GEAM. We alsopropose an on-line variational of GEAM model in this paper, which canprocess continuous Tweets, detect emerging events and track events evolution.We conducted experiments on over6million real tweets data to evaluatethe effectiveness of GEAM model. The experiments demonstrate that GEAMoutperforms the state-of-the-art topic model LDA in terms of Precision, Recalland DERate (measuring Duplicated Events Rate detected). Particularly,GEAM outputs a4-tuple (Time, Locations, Entities, Keywords) structure of thedetected events, which can provide fine grained information of the events.Wealso show that besides of effectively detecting events, GEAM can also be usedto analyze event trends. We then demonstrate the on-line GEAM model’sperformance on event detection in Twitter. Futhermore, the on-line GEAMmodel can track the evolution of events.