Research On Hierarchical Text Sentiment Classification Based On Comment Data

Recently,with the development of E-commerce,more and more users are expressing their opinions and comments on the Internet.Analyzing and distinguishing the sentiment tendency contained in these comments can create huge commercial and social value.Sentiment classification of user comments has become one of the research hotspots in the field of natural language processing.However,in the face of massive data sparsity and unbalanced sample distribution,manual processing can't meet the needs of traditional text sentiment classification method,which has low accuracy and granularity.In order to solve the above problems,this paper focuses on the hierarchical text sentiment classification,that is,mining the potential sentiment tendency of an attribute of an object.The problem is divided into two sub tasks: single label sentiment analysis and multi-label sentiment analysis.Firstly,the hierarchical text classification is designed and implemented,and the automatic category labeling system is constructed.Secondly,a recurrent convolution attention model is proposed for sentiment analysis.Finally,experiment is conducted on Sem Eval dataset.The research contents of this paper can be summarized as:(1)Research on hierarchical text classification method.In this paper,conducts experimental analysis on various models from two data sets of text classification sentiment analysis and compares the advantages and disadvantages of three mainstream model methods including CNN,LSTM and Attention.It is found that CNN is good at extracting local features.LSTM can model the long-distance dependence of text,and then encode the context information effectively.Attention can effectively integrate the features through weighting,which provides a sufficient foundation for the subsequent model fusion.(2)Multi-label Sentiment analysis based on LCA.Aiming at single label and multi-label problems in hierarchical text sentiment classification,a text sentiment analysis method based on LCA is proposed.LSTM is used to model the whole sequence and capture the long-term dependence.CNN is used to extract the local information in the context.Finally,the sentiment is predicted by softmax function.The experimental results on two single labeled datasets showed that the F1 is 82.1% which is equivalent to the state-of-the-art model.The experimental results on two multi-labeled datasets showed that the model is close to the baseline model on small datasets,and the F1 on big datasets exceeds the state-of-the-art model by 78.38%.The experimental results show that our method outperformed the traditional method in sentiment analysis,and can effectively improve the accuracy.(3)Construction of hierarchical sentiment analysis system.The results of LCA model are visualized and analyzed in this paper,and a hierarchical text sentiment classification system is constructed to demonstrate multiple labels from the perspective of visualization.