Research On Storyline Mining Based On Weibo

With the rapid development of social network,microblog like Twitter and Sina Weibo has gradually become a key platform for users to acquire and disseminate events,and especially a great number of posts revolves around newsworthy events.Therefore,there is an urgent demand for users to develop a Weibo-oriented storyline automatic mining method.The purpose of storyline mining is to analyze the development of events that some users are interested in,which provides an intuitive way for users to accurately understand the evolution of events.Compared to news articles,the storyline generation from microblogs is not an easy task.First,there are errors and nonstandard problems in the time information of microblog,which makes it difficult to normalize the occurrence time of events.Second,there is tremendous homogeneous and useless information in microblogs,which will worsen event extraction.Third,data sparseness and the lack of context make it challenging to learn the correlations among events when constructing story branches.In this paper,we aim to propose effective solutions to generate storyline from microblog posts automatically.And three aspects of research have been carried out to solve the problems of temporal expressions normalization,representative event extraction and events link:For the first issue(temporal expressions normalization),we propose a defuzzification algorithm based on sequence matching to infer the fuzzy time.Firstly,a rule-based temporal expression detection method is used to identify the timestamp of the event.Furthermore,the event evolution characteristics are used to analyze the time process,and the event occurrence time is obtained based on sequence matching algorithm.For the second issue(representative events extraction),we propose a new socialinfluence-based model with the temporal distribution to extract representative events from microblogs.Our model suggests modeling the social attribute of microblogs,analyzing the social influence of events,and then extract events with significant social-influence as representative events.Further,we propose a temporal distribution-based algorithm to improve the precision of event extraction.For the third issue(events link),we present Event Graph Convolutional Network(EGCN),an improved Graph Convolutional Network(GCN)model,to learn the latent relationships among events to link events and predict an event's story branch.We build a heterogeneous event graph to model the relationship between event nodes and events.Then,we propose a node-category prediction algorithm to link events.Further,we design a divergence-based method to solve the problem that GCN requires several nodes with labels as the input,avoiding human interference in the storyline's automatic construction.Finally,in order to evaluate the effectiveness of proposed method,comparative experiments and analysis are carried out in two real datasets with existing methods.The experimental results show that our method has better experimental performance in multiple dimensions.In addition,we build a prototype system of storyline mining based on microblog,which can automatically mine and generate the storyline of microblog events,and visualize the results of storyline mining.