Research And Application Of Text Emotion Analysis Method Based On Deep Learning

Changing the trend of gradual digitization of news media,an increasing number of news media use online channels to publish.Among them,there are not only giants in the traditional media industry,but also self-media that has grown up with the development of the Internet.Sorting and summarizing the news data of major news websites,analyzing and counting the relationship between emotional polarity and social sentiment reflected in current news hot news,has essential applications in media research,sociological research,and other fields.Considering the special requirements of the news industry for the objective and neutral content expression,it is difficult to judge the emotional changes expressed in the news text using the word embedding technology which is relatively weak in text expression ability.In this thesis,we designed a text emotion classification model based on the integration of Multi-head Attention and Long Short Term Memory(LSTM).By introducing multi-head Attention mechanism to update the Embedding coding method,the coding sequence was obtained through the full connection layer,which was then placed in the multi-head Attention algorithm module to weight the words in the text according to their importance.On the basis of this model,Bert model is introduced again for word embedding.The designed model preprocesses the extracted news text data simply,embedding words through the pre-trained Bert model,and further extracting features through the two-way LSTM network.The improved model was compared,fed back and trained with the word embedding representation of WWM-Bert pre-training model and Roberta model in the Chinese environment,and a reliable news text sentiment classification model was finally obtained.In order to evaluate the effectiveness of this model,this thesis selects the data set provided by the Internet News Sentiment Analysis Competition in CCF Big Data and Computational Intelligence Competition to verify the algorithm performance,and conducts comparative experiments with CNN,LSTM and Transformer,etc.Experimental results show that the Multi-head Attention-LSTM model has an average accuracy rate of 79% for sentiment classification of news texts.After adding the Bert language processing model for word embedding,the design of Roberta and bidirectional LSTM model found that the accuracy rate could be further improved to 83%,which is significant higher than other models used as comparative experiments.At the same time,based on the Roberta +Bi-LSTM model,a set of news and public opinion monitoring system is designed,The application scenario of the model discussed in this thesis is preliminarily designed.The system can be used to analyze the emotional polarity of news text after training,And in the system for summary and display,to help users grasp the emotional trend of news and public opinion.