Research And Application Of Online Public Opinion Guidance Effect Evaluation Based On Deep Learning

With the epidemic of COVID-19 at the beginning of 2020,local cities have successively entered a state of "silence",and citizens' daily life has shifted from offline to online.Internet users can access epidemic situation,policy release and social news through social media platforms,and express their opinions or emotions on news content.Due to the asymmetry of information access in the early days of the epidemic,panic and anxiety were more common.The negative network atmosphere leads to the complex and changeable situation of network public opinion and numerous hidden dangers.Finding out the hidden dangers of public opinion as soon as possible and respond positively,and correctly guide the public opinion through interpretation,disclosure,punishment and supervision methods,which can control the public opinion situation in time and avoid a wider range of adverse effects.The evaluation of the public opinion guidance effect can help managers to trace the events,accumulate effective guidance methods,find out the deficiencies in the work and timely adjust the strategies,and constantly improve the guidance plan.This thesis is mainly based on the microblog content from January to April 2020 crawled by the Internet,puts forward a variety of public opinion guidance effect evaluation indicators,and verifies the rationality and effectiveness of the evaluation indicators through theme evolution and sentiment analysis.The main work and innovations of this thesis are as follows:(1)In order to solve the problem of data annotation and avoid the subjective influence introduced by pure manual annotation,this thesis adopts the pseudo-label method to construct the sentiment multi-classification annotation data set of microblog news and comments.The pre-labeled model was trained using the SMP2020-EWECT dataset and used for microblog comment annotation,filtered for annotated comments that reached the prediction threshold and manually checked.(2)In this thesis,a cluster-based LDA theme model is constructed to extract the distribution of themes from daily microblog user releases and hot search,and then cluster all the themes to obtain the overall theme distribution during the epidemic period,and cluster the theme extraction results during the duration of public opinion,so as to realize the subdivision of the theme distribution of public opinion topics during the epidemic period.The distribution of subdivided themes can show the focus of netizens and the change process of attention more specifically,which is conducive to a more detailed and comprehensive analysis of the theme evolution work.(3)This thesis proposes a sentiment multi-classification model based on bidirectional attention mechanism to establish attention flow between microblog comments and news.The attention of comments to news perceives context,and the attention of news to comments introduces more semantic information for sentiment recognitions.Compared with the Bi GRUself-Attention,the accuracy of the model proposed in this thesis is improved by 3.61%.(4)This thesis puts forward a group of public opinion guidance measures evaluation indicators and a group of public opinion actual guidance effect evaluation indicators,and explains each indicators.The above indicators can evaluate the effectiveness and practical effect of public opinion guidance measures,help managers to trace the guidance process and guidance effect,and conducive to summarizing experience and improving the guidance scheme.