Research On Weibo Text Sentiment Analysis Based On Deep Learning

With the advancement of information technology and the rapid development of the Internet,more and more people have become Internet participants,expressing their views through the Internet and social platforms and sharing their knowledge.These contents contain comments on social hot issues,evaluations of goods and services,and so on,all of which have obvious emotional tendencies.Analyzing,processing,summarizing and reasoning these texts is of great significance for market research,user analysis,network public opinion discovery and early warning,and has important social significance and commercial value.Traditional research on sentiment analysis can be divided into approach based onsentiment dictionary and the approach based on machine learning.Approach based on sentiment dictionary usually require manual construction of special sentiment dictionaries for different fields.The quality of sentiment analysis is closely related to the quality and coverage of sentimental dictionary.The construction and maintenance of sentiment dictionary also requires a lot of manpower.With the emergence of new words on the Internet,approach based on sentiment dictionary can no longer meet the needs.The method based on machine learning relies on the manual selection of features.Different feature selections can cause differences in the results of sentiment analysis.There is a certain degree of difficulty in model generalization.In recent years,the method of sentiment analysis based on deep learning has become a hot topic of research.Compared with traditional machine learning methods,deep learning methods have stronger expressive capabilities and do not require manual feature selection and construction.They have a good effect on sentiment analysis tasks.In this thesis,based on summarizing the traditional sentiment analysis methods andthe existing deep learning-based sentiment analysis methods,the following types of deep learning models are proposed.Firstly,a CNN-BLSTM model which is combination of convolutional neural networksand bi-directional long-short term memory neural networks is proposed.Convolutional neural networks can extract hidden features from texts and combine features.Bi-directional long-short term memory neurons can use text temporal relations to learn sentence semantics better,and store context information while taking into account future context information.Experiments show that in the text sentiment analysis task,the performance of the CNN-BLSTM model proposed in this thesis is improved compared to other comparison models,including emotion dictionary based model,support vector machine model and other neural network models.Secondly,considering the situation that different words in the text are of different importance to the sentiment analysis task,the CNN-BLSTM-Attention model is proposed using the attention mechanism based on the CNN-BLSTM model.The attention mechanism can help the model obtain the semantic code containing the probability distribution of attention,and effectively highlight the more critical words in the text for the sentiment analysis task.Experiments have shown that using attentional mechanisms has higher accuracy in sentiment analysis tasks.In addition,this thesis also validates the influence of various parameters in the model proposed in this thesis on the final experimental results through comparison experiments,including four parameters:word vector quality,number of epoch,Dropout value and Batch size.Through experiments,it is found that the quality of the word vector has a positive correlation with the model effect.While the number of epoch is different,the model effect increases first and then decreases with the number of epoch.This thesis also found if the dropout value is too large will decrease the model effect,and the training time and model's effect should be considered when selecting the batch size.