Research On Opinion Mining In Public Comments Based On User Generated Content

Rapid development of the Internet has contributed to the widespread adoption of social network platforms.Chinese social network platforms represented by Weibo have attracted a large number of netizens with their free and convenient information interaction methods.News dissemination has shifted from newspapers,television,and other traditional modes to mobile internet positions.Netizens have a high degree of participation in network public opinion events,and the opinions and comments they publish not only reflect their attitudes towards the event,but also their sentiment tendencies and attitudes towards online media,especially dissatisfaction such as query and opposition,which affect the credibility of the media,destroys the media image,and then hinders the healthy development of the Internet.Therefore,what type of information the network media should provide,what type of image it should establish,and how to play a positive role in the dissemination of public opinion are significant issues to investigate.This paper constructs a research framework for opinion mining in public comments based on user generated content.Taking Weibo comments as the research object,the research on opinion mining in public comments is carried out from the two dimensions of sentiment analysis and stance detection.The purpose of this paper is to monitor the information experience perception attitude of netizens after the release of false news in the network media,and to provide a basis for the network media to improve the quality of content service,enhance media value,and for government public opinion guidance.The research is carried out from the following three aspects.(1)Based on the analysis of the composition of opinion mining tasks,this paper divides the research framework of opinion mining in public comments into three modules: corpus construction,sentiment analysis and stance detection.Considering the complexity of Weibo data and the needs of subsequent research and analysis,this paper proposes an efficient corpus preprocessing and construction method.Firstly,a preliminary corpus is constructed by crawling microblog comments and forwarding contents under relevant topics based on selected topics.Then use users as entities,comment and forwarding as the relationship to build public opinion dissemination user theme map,and find opinion leaders based on node centrality and user influence.Finally,the network media is selected as the target user from opinion leaders,and the user generated content of relevant microblog comments is crawled to construct the opinion mining task corpus.(2)In view of the problem that top-down autoregressive model cannot make full use of the context information,and the self-coded language model BERT ignores the dependence between the masked words,this paper proposes the sentiment classification models ERNIE-Text CNN and ERNIE-Text RNN＿Att based on the knowledge-enhanced language representation model ERNIE,and compares and evaluates the performance of the models through experiments and theoretical analysis.Experiments show that the three-level masking strategy of words,phrases and entities in ERNIE can obtain more semantic information.By inputting the word vectors trained by ERNIE into Text CNN and Text RNN＿Att models for training to obtain deeper semantic information,the effect of sentiment classification can be effectively improved.(3)Considering the influence of user sentiment and cognition on stance,this study extends stance detection from text classification to classification based on multi-dimensional feature fusion.This paper develops a comment stance detection model based on user feature fusion,which fuses user sentiment feature,user cognition feature and text feature.Text CNN is used to extract text features,and the fusion method is adopted in the feature layer to fuse user sentiment features and user cognitive features with text features for model training and stance prediction.The performance of the model is compared and evaluated through experiments and theoretical analysis.The experimental results show that the stance detection model based on user feature fusion performs better.