Hot Event Detection And Storyline Generation Using Microblog Data

With the rapid development and widespread use of internet technology and mobile devices,social media has already stepped into people's daily life,the information disseminated via social media has become an important source for acquiring news.Comparing with traditional media,users of social media are not only consumers but also creators and communicators,social media provides a novel way for information propagation,information exchange and information sharing.As a major form of social media,microblog holds hundred millions of users,huge amounts of user-generated content make microblog become an important data source for detecting and analyzing hot events.Due to the openness of microblog platform,information published through the platform is huge amount and redundant,excessive information will lead to information flooding unless we process it correctly.Thus,for the purpose of fast and efficient information acquiring,we should effectively detect events from microblog data,organize related information and make summaries of them.Based on this aim,we study event detection and storyline generation using microblog data,the main work is as follows:1)We propose an efficient approach for event detection from huge amounts of microblog data,and we train a binary classifier based on three different features,it classify the event into real or virtual event.Experiment results indicate effectiveness of both two methods.2)We build a multi-layered event characterization model to encode the relationship of posts from four perspectives(time,text,image and user interaction),then we fuse the four layers together to generate a storyline to summarize the evolution of the event.Experiment results indicate that our method provides fine-grained event summaries(dynamic evolution)and effectively identifies strong correlations of clues.3)The multimedia data in microblog is rich but redundant,thus,it is necessary to select a representative set of images for the event summary via calculating the correlation between the text and the image.However,the presentation space of text and image is not the same,therefore,we propose algorithm CMM based on cross-media data association.Experiment results indicate that comparing with the previous methods,CMM makes a balance between relevancy and diversity on image selection.