Research Of Structure Based Document Level Sentiment Analysis Method

With the continuous development of the Internet,sentiment analysis,as a means of mining user opinion information,has quickly become a hot research direction in the field of Natural Language Processing.However,analysis of the document level sentiment still has many challenges to be overcome.On one hand,there may be the opposite polarities in the long text.On the other hand,there may also be a large part of the text have nothing to do with the writer's attitude,which affect the emotion the results of the analysis.In view of the above problems,this paper proposes a document level sentiment analysis method based on rhetorical structure parsing and a document level sentiment analysis method based on opinion and non opinion part separation.In rhetorical structure analytical method,the document is parsed into a rhetorical structure of the parse tree,then we use the dictionary based method and method based on recurrent neural network to calculate the discourse unit emotional value,finally by using the rulebased method,we sum up the discourse unit emotion value along the rhetorical structure of the parse tree bottom-up recursively.In the opinion and non opinion part of the separation of the document level sentiment analysis method,this article uses manual annotated opinion and non opinion datasets to train classifier,and use this classifier to separate the opinions from non opinion in training set and the testing set of sentiment classifier.Experiments show that the methods proposed in this paper have greatly improved in accuracy,which proves the validity of the methods.In addition,the rhetorical structure parsing based analysis method has good adaption of domain of dataset and the method based on opinion and non opinion separating has good generalization ability.