Research On Microblog Rumor Early Detection Considering Emotion Analysis

With the gradual development of Internet technology,social media platforms have become an indispensable part of peoples life.Users have become accustomed to express their opinions on them,and the amount of information on the Internet has increased rapidly.However,some problems also follow,the quality of information on the Internet is uneven,and filled with all kinds of rumors,seriously affecting peoples access to information,if not dealt with in time,will even endanger social harmony and stability,so the automatic detection of rumors in social media has become an important research content.The existing rumor detection has the problem of low timeliness.Usually,rumor events in the later stage of transmission are detected.Due to the lack of relevant clues,prevention cannot be carried out before the rumor causes great harm.To solve this problem,this paper selects Weibo,the most widely used social media platform in China,as the research object.As far as possible,more relevant features are mined from micro-blog rumors,and model is carried out based on deep learning method to realize the early detection and prevention of rumors.The main work of this paper is as follows:1.A convolutional neural network model combining the attention mechanism of bidirectional affective words(DSA-CNN)is proposed,in order to fully extract UGC text emotional features.The attention mechanism is introduced into the pooling layer of convolutional neural network,and the existing emotion dictionary is extended to construct a mixed emotion dictionary for social media.Two kinds of pre-training methods of sentiment word vectors are proposed.The positive and negative sentiment word vectors are obtained by mixing sentiment dictionary and Word2Vec.As the attention pooling query vector,the final emotion representation is obtained after pooling and merging.Bidirectional attention pooling can extract more features from short texts from both positive and negative emotions.Experimental results show that compared with the classical classification model,the DSA-CNN model has better performance on two social media sentiment datasets,and the convergence speed of the model is significantly improved.2.A rumor early detection modelbased on multi-feature(REDM)is proposed to realize early automatic rumor detection on weibo.Firstly,the feature index system of rumor events is constructed,including external feature F_o,semantic feature F_s and affective feature F_e of rumor corpus.Secondly,according to the users historical Posting record,three behavioral characteristics of historical emotional tendency,emotional fluctuation and credit degree are calculated.Next,BERT model and DSA-CNN model are used to extract semantic and emotional features of rumor corpus.Finally,considering the temporal characteristics of event propagation,LSTM model is used for sequence modeling.The extracted semantic features,emotional features and external features were combined and input into LSTM to detect rumors.According to the extracted characteristics of users historical credit,microblog users are divided into immune users,vulnerable users and bad users.The statistical analysis of the basic and behavioral characteristics proves that there are certain differences in the attributes of different user groups,which can effectively assist rumor detection.Experimental results of rumor detection on Weibo dataset show that the proposed model has the best detection accuracy compared with other baseline models.The first several posts of the rumor event were intercepted as detection nodes,which proved that the model still had good performance in detecting the rumor in the early stage.