Research On Micro-blog Sentiment Classification Based On Recurrent Neural Network

As a social networking platform which develops rapidly these years,micro-blog due to its easy operation,fast spread,high flexibility and other characteristics,has been widely advocated and used by users.Although micro-blog text contents posted by users are very complicated,they hide a lot of useful information by observing and analyzing them,especially the sentiment orientation in the text,which can help the government and enterprises to understand the needs of the public,guide the public opinion,find business opportunities or increase revenue.So far,more and more scholars have paid close attention to the research of sentiment classification of microblog text.How to learn deep semantics,express text features effectively and improve the effect of sentiment classification have always been the research goals in related areas.This paper mainly studied two aspects of micro-blog text sentiment classification: subjective and objective classification,and sentiment polarity classification.In the stage of subjective and objective classification,a method based on the dictionary and corpus was proposed.In the sentiment polarity classification stage,the feature extraction methods and classification algorithms of micro-blog text were studied respectively.For the feature extraction,a feature fusion method based on shallow and deep learning was proposed.For the classification algorithm,a sentiment classification method based on a kind of improved recurrent neural network was proposed.The main work and innovations of this paper were as follows:(1)For the subjective and objective classification of micro-blog text,a method based on the dictionary and corpus was proposed.Firstly,the reliable subjective text was identified according to the reliable sentiment dictionary,which was constructed by this paper.Secondly,the remaining text was classified as subjective text or objective text,combined with the method of corpus statistics.Finally,the F1 value was 6.72% higher than the traditional subjective and objective classification method based on the massive sentiment dictionary.(2)In view of the general shallow learning features ignoring the intrinsic semantics of the text,a feature fusion method based on shallow and deep learning was proposed.The shallow learning features included three kinds: unigram,part of speech and dictionary.Deep learning features were extracted by the word2 vec tool.Then they were fused by using two schemes.The experimental result showed that the sentiment polarity classification effect of micro-blog text with feature fusion was better than those with only one kind of features.(3)For the sentiment polarity classification of micro-blog text,an improved recurrent neural network model was adopted.The model replaced the hidden layer of general recurrent neural network into LSTM structure,which not only took the correlation of the text sequence into consideration,but also could learn the distant information in the text.The final classification accuracy rate was 85.04%,which was 3.17% higher than that of the traditional SVM with shallow learning features.