The Research On Sentiment Classification Based On The Deep Learning Models For Text Data

With the rapid development of the Internet and information industry,a large number of text messages,such as views,opinions and comments containing sentimental tendencies,have emerged on Weibo,Twitter,and other online platforms.Mining the emotional polarity(positive or negative)contained in these text messages is of valuable and important to commodity purchase,product optimization of businesses and public opinion understanding for the government in real time.Existing research methods face the great challenges due to the existence of turning sentences and the lack of label data.Based on the deep learning model,this dissertation focuses on the sentiment classification of text data and two main challenges mentioned above,and it has important research and application value.The main research method in this dissertation is built on the model of deep learning.Main contributions are shown as follows:(1)This dissertation summarized the related work in sentiment classification,cross-domain sentiment classification and deep learning,including the definition,related theory,research background,and research status.(2)For the existing sentiment classification methods,it is difficult to predict the sentiment polarity of sentences with turning semantics accurately.A segmented pooled convolutional neural network model(PPCNN)is proposed in this dissertation.This method segmentes the convolution results based on the position of the transition word in the sentence,and then uses the maximum pooling to extract important features of the local segment.Finally,the Dropout method and the softmax classifier are used in the sentiment classification.This method can extract important features of multiple polarities and preserve the relative order of features.A large amount of experiments demonstrate the effectiveness of this method.(3)This dissertation,we proposed a cross-domain sentiment classification method based on the model of word2vec(WEEF),which aims at the problem of expensive labeled information of massive data in sentiment classification.Our method selectes high-quality domain co-occurrence features as a bridge,and uses these features as seeds.Based on the similarity calculation of word vectors,domain-specific features are extended into these seeds to form feature clusters,and the differences between domains are reduced.Experimental results demonstrate the effectiveness of the method.