Research Of Users Attitudes Analysis Towards Hot Topics In Social Networks

Social networks have continued to develop in recent years,and the user scale has also gradually grown.Social networks provide a platform for interaction and communication where users can participate in discussions on hot topics and express their attitudes.Analyzing users attitudes towards hot topics in social networks has important research significance and application value.However,as to the complex topological structure of social relationships in social networks and the unstructured text data posted by a large number of users,it is urgent to find a method to automatically analyze users attitudes towards hot topics.In this thesis,we collect 3 hot topic datasets from Sina Weibo,and divided the users into posted users and non-posted users according to whether the user has posted tweets related to the topic,and then use graphbased semi-supervised learning to classify the attitude polarity of these two types of users.As for the task of analyzing the posted users attitudes,we propose a method based on text encoder+GNNs,in which the text encoder is used to extract attitude features from text while GNNs are used for semisupervised learning on the graph of social relationship among users.In the text encoding stage,the influence of part-of-speech on attitude features is specifically studied,and a part-of-speech information gain gating algorithm is proposed.The algorithm enhances the ability of the text encoder to extract attitude features from text by introducing part-of-speech embeddings and information gain values as the global prior knowledge of the model;in the semi-supervised learning stage,we explore the impact of the users influence on the spread of attitudes in social networks,and propose a social relationship direction gating algorithm.The algorithm can be flexibly merged with existing GNNs,learn the flow of attitude information between nodes more effectively,and expand the path of label propagation.As for the task of predicting non-posted users attitude,we propose an end-to-end attitude analysis framework.It integrates the two tasks that judge the attitude polarity of posted and non-posted users into an end-to-end semi-supervised learning framework based on text encoder+GNNs.Experimental results on the 3 hot topic datasets and NLPCC stance detection shared task public datasets show that the part-of-speech information gain gating algorithm,social relationship direction gating algorithm,and attitude analysis end-to-end framework proposed in this thesis can further improve the accuracy as well as the macro F1 score of attitude classification.Finally,this thesis implements an demonstration system for users attitudes analysis towards hot topics in social networks.