Research On Public Opinion Theme And Sentiment Evolution Based On BERT And HDP-vMF Mixture Model

The Internet provides a more open and equal communication platform for the public,while enabling the public to access and share first-hand information at any time.However,the Internet is a double-edged sword,which will not only trigger a large-scale public opinion crisis,but also become a relief valve for social emotions.In the Internet era,the network has gradually become one of the main media of information dissemination.The speed of information dissemination through the network is faster and faster,and the piblic participates in topic discussion through the network more and more frequently.Among them,in the process of network communication,public emergencies are more likely to evolve into public opinion crisis.If the media and the government find problems in time,clarify rumors and dredge the public’s negative emotions,they can be contained in the bud.In view of this,in order to help the government and media managers carry out the supervision of online public opinion more efficiently,this paper establishes a hybrid model.While mining the theme of comment text,the model explores the emotional trend of Internet users in the whole stage of public opinion,so as to improve the comprehensiveness and accuracy of public opinion analysis.In terms of topic mining,the model uses the hierarchical Dirichlet mixed process model(HDP VMF)to fine-grained extract the hot topics of microblog comments.HDP VMF regards each comment data as an independent group,extracts fine-grained topics based on group data clustering method,and replaces Gaussian distribution with von Mises Fisher distribution,so that all clusters are evenly mapped to a unit hypersphere,rather than encouraging cluster centers to converge to the original point,so as to reduce the error in the clustering process.In the aspect of emotion evolution analysis,the deep transfer learning model(BERT)is used to model the fine-grained emotion classification of microblog comment data.BERT,which is composed of attention network,significantly improves the running speed by using parallel operation.At the same time,the multi head attention mechanism also enhances the ability of semantic analysis and understanding of the model,and provides a more matching theoretical algorithm for the research of critical emotion analysis.In terms of empirical analysis,taking the "rainstorm in Henan" event as an example,this paper captures the data of microblog comments on relevant topics from July 19 to July 31,2021,with a total of 120791.After the preliminary cleaning of the data,the jieba tool is used to segment the data,and the word2 vec combined with BERT is used to convert the segmented data into word vectors.Then,the sentence vector is obtained by weighted average word vector,and the sentiment classification of comment data is realized by BERT,and based on this,the sentiment evolution diagram is drawn.Finally,with the help of HDP-v MF model,the key words of the review data are extracted,and the hot topics concerned by users in Weibo are obtained.The empirical results show that:(1)Emotion evolution: In the period of public opinion outbreak,most of the negative emotions of Weibo users,but the overall emotional tendency in the whole cycle is positive.(2)Hot topics: "rainstorm in Henan" is the core topic discussed by microblog users during the period of public opinion.With the further development of the incident,a number of hot topics have emerged,among which "Metro Line 4 flooded","star and enterprise network donations","k599 train trapped" and "rescue documents" are the top four topics,HDP-v MF can effectively mine real topics through verification with Weibo hot search topics.(3)BERT performance: through the performance verification of BERT and six commonly used emotion classification algorithms in academic circles,BERT’s ACC and F1 values rank first,which verifies BERT’s superior performance in emotion classification.