Research On Detecting Methods On Micro-blog Rumors

At present,sina microblogging is one of the most popular service provider and information dissemination base with the largest number of active users.It has a numerous information,and propagates so freely,so it owns huge influence,and gradually becomes an important medium for public opinion erupting and warming.The rumor auto-detection research has attracted considerable attention in recent years in natural language process and data mining domain,in order to advance warning,prevent,monitor,manage and other rumor clean works.There are a large number of related studies on rumor detection based on twitter and sina microblogging.Most existing rumor detection methods typically exploit supervised machine learning methods based on a wide range of feature engineering corresponding to user aspect,texts aspect,social networks aspect.However,traditional research needs linguistic and psychological background.Moreover,the feature extracting process is complex,and is prone to feature redundancy,featuremissing phenomenon.Deep learning models can refine and abstract text,extract feature automatically,even to an infinite space.Meanwhile,feature learning is integrating into model construction,without manual design features.This paper mainly studies the application of GRU and LSTM model in rumor detection research,to judge whether a microblogging text is rumor information or not.Considering the graph structure of sina microblogging,that is to say,a microblogging text correspond to a number of comments,and comments may contain attitude towards this text,such as favor,oppose,doubt.In addition,comments are post by multiple users,contain rich opinions or suggestions.Therefore,In our research,we use two different methods to model the comments,learn semantic representation,weigh the relation between microblogging text and comments.The first model is comments representation learning based on attention model,which regards comments as time nodes on a time line.After unfold by time,every comment can be treated as the input of time series model at every time.We also propose attention mechanism,which weighs importance degree of every time node to final semantic representation.The other model is comments representation learning based on memory network which considers that every comment is independent of each other.And each comment is placed in slot of the memory array.This model selects comments which have high correlation with input sentence,and apply the comments in the calculation and construct classifier.