Research On Sentiment Analysis Based On Deep Feature Representation

With development of social media like Microblog,BBS,Zhihu and Douban,more and more people tend to express opinions,share knowledge and make original works rather than just browse information through those social media,and all these become a part of Internet resources.These content contains many comment of hot issues or a specific commodity,which have emotional tendencies.The mission of emotion analysis is how to mine valid information from these massive and unstructured data.Sentiment analysis had great commercial value,which is widely used in forecasting stock price,product analysis,and promotion and so on.For governmental agencies,emotional analysis is basic technology for government in public opinion monitoring,public opinion polls and crisis management.Traditional text emotional analysis depends on complex feature engineering,which take a lot of manpower in feature selection,and the selection work may restart if application area changed,so the traditional way is not universal.From this year,deep learning is widely used in natural language processing,so this paper summarizes existing deep learning model which can be used in emotional analysis and establish two deep learning model on the basis of summary of traditional emotional analysis methods.The first is RSGRU network which combines Recursive Neural Networks(RSNN)with Gated Recurrent Units used in sentence level sentiment analysis.Compared with widely used Long-Short Term Memory(LSTM)and convolutional neural network,applying this networks in sentence level emotional analysis mission make better performance of text emotional polarity classification mission,and compared with LSRM,RSGRU Networks have less parameter and is easier to train.The second is CNN-BGRU model which combined convolutional neural network and bidirectional gated recurrent units,we apply this model in sheet level text emotional analysis.Convolutional neural network have a better performance in extracting local feature.Bidirectional gated recurrent unit is more practical in expressing sequence information.This paper with combination with this two model get better result in emotional polarity analysis,and compared with traditional way,and there is no need man-made feature extracting,which can save a lot of manpower and is easier to maintain.