Microblog Sentiment Analysis Based On Multi-features Fusion

With the advent of the Internet age and the progress of science and technology,the social network platform represented by Weibo has gradually become the main way for people to share and obtain information.The information posted by users on the social network platform includes discussion of social hot events,evaluation of commodity services,etc.,and mostly with obvious emotional colors.Sentiment analysis of these subjective information helps promote the research of public opinion analysis,personalized recommendation and emergency prevention,etc..Therefore,it is of great values to study Weibo sentiment analysis.This paper focuses on the sentiment analysis of Weibo texts and images,the main work includes:(1)Previous studies less focus on some factors such as user personality and content characteristics,and this may lead to unsatisfactory Weibo sentiment analysis.To address the issues,approached a text sentiment analysis method,which fusing user-based features and content-based features.A text sentiment classification model,TSCCUF was proposed,which considering the user-based features and content-based features have a good indication of emotions.The experimental results show that the fusing of user-based features and content-based features could enhance model ability of capturing emotion semantic,and compared to BLSTM and MCNN models without feature fusion,the performance of TSCCUF was greatly improved.(2)To solve the problem of overfitting or slow convergence for image sentiment classification model constructed with CNN.A image model based on parameter transfer and fine-tuned was constructed.And to solve the problem of the accuracy of sentiment analysis for merely texts or images cannot be greatly improved.A multimedia approach for Weibo sentiment analysis was proposed.Based on the TSCCUF and TFCNN,the early fusion and late fusion methods were designed respectively for visual-textual sentiment classification.The experimental results show that joint visual-textual sentiment analysis could further improve the accuracy of Weibo sentiment classification,and better classification results could be obtained by late fusion.(3)Most Weibo sentiment analysis methods less focus on emotions dependency between Weibos,to address the issues,approached a Weibo sentiment analysis method,which fusing conversation-based features and forward-based features.A Weibo sentiment classification model,TSCCFF is proposed,which considering conversation-based features,forward-based features and Weibo sentences to make up for the sparsity of data,and captures the emotional dependency between Weibo sentences and conversation-based features and forward-based features.The experimental results show that TSCCFF outperforms the typical approaches including CNN,LSTM and attention-based Bi-directional LSTM.