Research On Event Detection For Short Text Of Weibo

Events that happen in real-world and users' opinion over a wide range of topics spread in the micro-blog services platform such as Tweeter, Sina Weibo. The events can be variety, such as celebrities or political affairs, socail local events and natural disasters. Messages are posted by users after they have experienced or witnessed the events happening in the real-world and they want to share their experiences immediately.The user data and text data is a enourmous valuable resources for research such events detection in socail network, however, There is two challenge for event detecting in soical media:first, the social media texts are unstructured and full of noise, for which many existing algorithms cannot apply. Second, the traditional based text clustering algorithms will become very unefficient as the volumn of social media dataset increase.We propose two methods for event detection in social media in this paper: Event-Graph based event detection and EventLDA, a bayesian generative model.In EventGraph model, we build a weighted directed graph, EventGraph whose nodes are words and exits an edge between nodes if the nodes co-occurrence in a doc-ument. The relationship between nodes in EventGraph is able to capture the semantics among them. The events are embeded in the EventGrahp in the form of sub-graph or communities. We propose a key node based event community detection method, which improve the efficiency of graph based event detection algorithms. Besides, it's very easy to parallize the algorithm since key node based communities extraction are inde-pendent.EventLDA is a three layer bayesian generative model, which includes words, events and user. In this model, we model events as a mixture distribution of words and user as a distribution of events. In this model, the production of a microblog is assumed as :the user first choose a event based on its event distribution. Then she chooses a bag of words one by one based on the chosen event until the microblog is done. The objective of this model is to get the events' distribution of the words, which is estimated by Gibbs sample method.In order to evalute the proposed methods, we conducted experiments on three pub-lic datasets and one sina weibo dataset. The results showed that our algorithms are more efficient than the LDA and the BNgram. Besides, we explore how the parameters in-fluence the result.