Research On Chinese Text Sentiment Classification Method Based Of Deep Learning

With the continuous development of the Internet to intelligence,digitization and mobilization,we not only passively receive online information that was always done in the past,but also are willing to create online data recently through blogs,network discussion groups,business communities and other platforms.The large amount of unstructured textual information that is generated by the Internet everyday contains the user's views,attitudes,and emotions about people,things,and events.Such information has great value in industrial and academic research field.If the emotional information can be extracted from these massive texts,it will benefit the development of Internet public opinion analysis,business management,event forecasting,and merchandising.In the field of text sentiment classification,traditional methods based on sentiment lexicon and machine learning are only suitable for these cases where the amount of text corpora is small and the text semantics are simple.However,with the rapid increase of the volume of text corpus and the continuous enrichment of expression methods,the traditional classification methods are incapable of handling them.As a result,the technology of using deep learning to classify emotions has emerged and developed rapidly.This thesis applies Convolutional Neural Network(CNN)model of deep learning to the sentiment task of Chinese texts.Among the existing text sentiment classification methods,text expressions often use static word vectors,and they also fail to take into account the aspect of the multi-part feature information of the sentences.To tackle the problem of multi-part feature information,a segmented pooling and dual-channel CNN sentiment classification model was proposed,and significant improvements were achieved.The main contributions are as follows:(1)Glove model is used for text representation and it is optimized by word vector fine-tuning strategy.Such model is used to train the word vectors,and the back propagation of neural network is used to fine-tune the initial word vectors of the input network.By doing so,the word vector of the input network can be optimized continuously with the training of the CNN,which helps to improve the overall classification accuracy of the model.(2)A segmented pooled convolutional neural network(SCNN)text sentiment classification model is proposed.The existing CNN sentiment classification model failsto take into account the features and position information of multiple parts of a sentence,thus losing some sentence information.This thesis proposes a SCNN text sentiment classification model based on the segmentation pooling strategy.It fully considers the characteristic information of different parts of the sentence.In the pooling phase,the sentences are segmented and the eigenvalues are obtained respectively to improve the accuracy of the model.Through segmented pooling operation in our experiments,we observe that the classification accuracy rate of the optimized CNN can be improved by up to 4.3%,compared with traditional CNN model.(3)The dual-channel convolutional neural network(DSCNN)text sentiment classification model is proposed.With dual-channel text representation,more text feature information can be obtained while optimizing the text representation of the input network to improve the performance of the model.Evaluation shows that the glove +glove dual-channel model achieves the optimal results.Compared with the previous convolutional neural network sentiment classification model,the accuracy rate is improved by up to 5.8%.