Research On Sentiment Analysis Of Social Text With Capsule Networks

With the rapid development of society,there are millions of users,social medias such as We Chat,Taobao,and Weibo.Through social medias,they can express their moods,feelings and opinions on various events,and thus generate a large of social text data.By mining the emotions in social texts,the emotional tendency of users and social public opinion problems can be reflected intuitively.It plays an important role in the government's supervision of public opinion,in the management and decision-making of enterprises and in the management of personal emotions.At present,the analysis and research on the sentiment tendency of traditional texts has been relatively mature,but the sentiment analysis research of social short texts is still relatively backward.The sentiment analysis of social short texts mainly has the problems of irregular structure,sparse features and unsatisfactory classification effect.In response to these three issues,this article puts forward three corresponding research contents.(1)The pursuit of simplicity,randomness and openness in social medias have led to more and more mixed use of Chinese and foreign languages,causing irregularities in the structure of social texts.Therefore,this paper proposes to use the Chinese and English hybrid word segmentation technology based on string matching to reasonably segment the characteristic words of Chinese,Chinese and English,English and numbers,Chinese and numbers.In addition,in recent years,there have been more and more types of emoticons.These emoticons can not only fill semantic information,but alse can even supplement or change the emotional tendency of social texts.Therefore,in this paper,emojis are processed into text and then integrated into the text for sentiment analysis,which not only expands the text features,but also enriches the content and sentiment of social texts.(2)Because users are accustomed to publishing short texts to express their views on news events,their likes and dislikes to public figures,and their evaluations of products,or to vent their personal feelings.The resulting social text content is usually relatively short,causing feature sparseness.Aiming at the problem of sparse feature words in social texts,this paper proposes a feature correlation library expansion scheme based on information gain.Through the expansion of social texts,more features can be obtained for sentiment analysis tasks.The experimental results show that the accuracy of sentiment analysis experiment after feature expansion is significantly improved compared with that before expansion.(3)Due to the structural dependency of social text,a single neural network model cannot achieve contextual semantic extraction and local feature extraction at the same time.Therefore,this paper proposes a Caps Net-LSTM hybrid model that integrates emotional information.The model first uses Capsule Network to extract local semantics and location features,and then uses Long-Short Term Memory perform contextual semantic extraction,and finally use the softmax classifier to perform sentiment analysis of social text.The experimental results show that the accuracy of the Caps Net-LSTM hybrid model proposed in this paper is 4% higher than the LSTM model and 3% higher than the Caps Net model in the fine-grained sentiment analysis experiment.