Multi-label Public Sentiment Analysis Of Hot Topics In Social Networks

People spread real-time information on social networks through mobile devices,and then become the initiators and communicators of hot topics.Hot events propagate and spread on social networks,generating public opinion related to hot events,which in turn affects our real life.The analysis of public opinion in the network is of great significance.Public emotional information is an important part of public opinion.Current public opinion analysis is mostly from the perspective of topic evolution,ignoring the information of "emotion" in public opinion.From the perspective of "emotions",this thesis will show the development and change of public emotions in different time periods under the hot events in social networks through the multilabel sentiment classification method for the text content about hot events posted by users.Furthermore,it supplements the comprehensiveness and completeness of the existing public opinion analysis content.The specific work of this thesis is as follows:1)Considering the problem of unknown label sequence when using seq2 seq model to process multi-label tasks,this thesis proposes a dynamic label sorting algorithm based on conditional probability.The algorithm can dynamically adjust the label sequence in combination with the model prediction result set and the sample real label set to improve the model effect.Based on the sequence generation model SGM,an ASGM model incorporating the algorithm is designed.By comparing the training effect of the model on the RCV1 data set,the effectiveness of the algorithm is verified.2)For the neural network multi-label model such as ASGM,the model is difficult to train due to the huge parameter amount and the gradient disappears.The multi-label classification model TML-HN based on Highway Network is designed to effectively reduce the parameter amount and alleviate the gradient disappearance problem.Finally,the validity of the TML-HN model is verified on the RCV1 data set by comparison with other models and algorithms.3)Public sentiment analysis algorithms for hot events is designed.Models are trained by multi-label sentiment datasets,on the basis of verifying the effectiveness of the above two models.Then,multi-label emotion models predict the emotional tags of social network hot topic data.According to different time granularity,the process of public emotion development and change with time is shown through statistical.The prototype of a multi-label public sentiment analysis system that integrates three functions of crawler,text sentiment multi-label prediction and hot-spot public sentiment analysis is designed.