Research And Implementation Of Group Mining Technology Based On Topic Stance

With the rapid development of Internet,social network Apps have become the largest public opinion fermentation center,generating a large amount of data reflecting the tendency of social public opinion.These data have clear text stance,discrete network organization structure and sparse content features.Existing public opinion-related research generally focuses on text sentiment and network structure or user attributes.Starting from the stance group,this paper conducts three researches on hot topic extraction,stance detection and group mining to present the whole picture of public opinion.The main research results are as follows:First,research on hot topic extraction methods.The existing related solutions either complicated operation process,separate clustering and topic word extraction tasks,or complex solutions that cannot be calculated in parallel.This paper proposes a clustering method based on neural topic model,based on the variational autoencoder,combined with word2vec,the topic words extraction and clustering task are jointly modeled in an unsupervised form,which can output article categories and topics at the same time.Compared with its base model,the performance of topic words extraction is improved significantly,and the text embeddings of the same class are more aggregated.Second,the research of stance detection method.Social short texts Stance Detection are faced with many challenges,such as sparse features,new words,no target topic in the text,irrelevant stance and sentiment,and diverse and changeable topics.This paper designs a stance detection model based on co-attention and consistency constraint learning.Firstly,it uses Ro BERTa to obtain text embedding,then the co-attention mechanism is used to enhance the connection between the content and the target topic sequence,and finally uses consistency constraints to make model Self-learning the essential feature applicable to all topic stance classification.This model only uses one classifier to predict multiple topics at the same time,and have achieved state-of-the-art results on the NLPCC2016 Stance Detection in Chinese Microblogs with F_avg equal 78.69%,which was nearly 6%higher than the original optimal model.Third,research on group mining methods.In this paper,the group mining problem on social networks is abstractly modeled as a multi-attribute view node clustering problem.In order to deal with the complex and diverse network structure,this paper designs a deep clustering model based on multi-attribute view.Firstly,the feature graph is constructed according to the nodes with similar content,which is input together with the topology graph.Through the three-layer attention mechanism,the topology and the content features of nodes under different attribute views are adaptively extracted,and supervise and guide the node embedding learning from three aspects:the direction of embedding learning,the degree of embedding distortion,and the degree of interference of the clustering results on the embedding space.Comprehensive experiments show that the embeddings learned by this method are more suitable for clustering tasks,and can deal with nodes with sparse relationships.The state-of-the-art results are obtained on the four open source datasets of multi-attribute and single-attribute views at the same time,Compared with the best baseline on each dataset,the accuracy is improved by an average of 6%.Finally,based on the above three core algorithms,a prototype system of Group Mining based on Topic Stance was designed and implemented.The system can extract and display daily hot topics,generate topic evolution context,automatically identify and statistic the proportion of netizens'stance tendency towards hot topics,and display the results of group mining with the same stance,so as to assist in the control of public opinion.