Sentiment Classification Research Based On Semi-Supervised Recursive Auto Encoders

Text sentiment analysis occupies a pivotal position in data mining research, and text sentiment classification plays an important role in text sentiment analysis. Text sentiment classification uses some natural language processing methods to obtain and analyze sentiment of the text, including extracting all the elements related to sentiment of the text, and to judge the emotion tendentiousness of text, etc. The emotional orientation of text mainly includes positive, negative and neutral and so on.With massive comment texts on the Internet, how to analysis and utilize them quickly and efficiently caused the attention of more and more scholars. Now, researches on text sentiment are based on machine learning and emotional dictionary, both having got good results. But due to the method based on emotional dictionary mainly dependents on the built of dictionary, which is completely finished by people’s subjective experience, so many scholars focus on building a more complete dictionary. For the traditional machine learning method, in the process of modeling, simple function and simple methods are usually used, so they are considered as shallow study.In this paper, we introduce the semi-supervised recursive auto-encoder method to research text sentiment classification, which is a deep learning method. It studies the nonlinear network structure in deeply, and expresses characteristics of the text through distributed vector, and makes up constrains of shallow learning on express ability and generalization ability.In order to further analysis, we first establish a traditional SVM classification model for text classification, get the optimal result 87%. And with the semi-supervised RAE method based on the deep learning, when feature selection and pre-process are exactly the same as the case of SVM, the optimal result is 88.3%, increased by 1.3% than the traditional SVM methods.