Research On Text Sentiment Classification Based On Deep Learning

Text sentiment classification occupies a pivotal position in sentimentanalysis research. In the twenty-first century, with explosion of information,sentiment classification research on massive amounts of data has attracted a lotof researchers. How to learn the text semantic information deeply, how to expresssemantic features accurately, and how to improve the accuracy, are the targets ofsentiment analysis.Since the traditional machine learning methods have the disadvantage thatthey cannot acquire the text semantic information, this paper proposed themethod which integrates the deep learning features and shallow learning featuresfor text sentiment classification task. It can introduce the expression of semanticinformation, and improve the model’s ability of learning and understanding themeaning of texts. In experiments, the deep learning features are not extracted bymultiple hidden layers deep learning methods, which makes that feature vectorscannot really understand the specific semantic of the texts. To solve theseproblems, this paper introduced the semi-supervised recursive auto-encodermethod (RAE) to research text sentiment classification, which belongs to deeplearning. The structure of RAE is multiple hidden layers neural networks, whichcan analysis and optimize each feature expression layer by layer, then the finalfeature expression can express semantic information of text more accurately, sothe RAE method can improve classification performance.At first, our paper used the traditional SVM method to research the sentimentclassification task, in which we select the features combined of words, part ofspeech and dictionaries, and get the best accuracy of81.88%. In feature fusionmethod, the first step is to get the best size of deep learning feature vectors whichis150in an experiment. With the result, the feature fusion method achieves theresult of81.98%, which is improved by0.1%than the traditional SVM. In semi-supervised RAE method, the optimal value of RAE feature vector is50. It has asignificant improvement in accuracy, that is85.10%,3.2%more than thetraditional SVM. When the size of corpus was enlarged by double, the accuracyraised2.5%, with two more times of learning time cost than original.