Research On Text Sentiment Classification Based On Deep Neural Network

With the development of social platforms and e-commerce,internet users are experiencing an explosive growth trend,generating a large amount of text,audio,pictures and video data every day,while the amount of text information is large and chaotic,and it is difficult to distinguish and organize manually.The massive amount of text data generated on the internet reflects the changes in users' opinions and emotions.It is more and more important to mine the emotional information contained in massive text data reasonably and effectively.Therefore,with the development of Word Embedding to train text data to low-dimensional dense vector,combined with deep neural network,it has gradually become the main technology of text sentiment calssification.The main research content of this paper is to combine the convolutional neural network and the recurrent neural network to complete the text sentiment classification task.Convolutional neural networks have certain advantages in extracting text features through convolution operations.However,convolutional neural network models tend to ignore the contextual semantic information of words when performing sentiment analysis,resulting in polysemy.Therefore,this paper proposes the C_BiLSTM(CNN_BiLSTM)model,which combines the local features extracted by the convolutional neural network and the global features extracted by the Bidirectional Long Short-Term Memory.BiLSTM(Bidirectional Long Short-Term Memory,BiLSTM)overcomes the gradient disappearance or gradient explosion problem of traditional recurrent neural networks.BiLSTM processes long-distance contexts from forward and backward directions,extracts global features with context information,to a certain extent,it can avoid the problem of polysemy in CNN for text analysis.The model can not only take advantage of CNN to extract local features of text,but also use BiLSTM to extract global features of text.Experiments on SogouCA dataset show that the proposed model achieves good results in the text sentiment classification task.The C_BiLSTM model proposed in this paper solves the problem of polysemy to a certain extent,and achieved a good text sentiment classification results.However,the convolution neural network of C_BiLSTM model has a shallow number of layers and can not effectively extract high-level text features.So this paper proposes the VDC_DBLSTM(VDCNN_DBLSTM)model,which uses 8 convolutional VDCNN(Very Deep Convolution Neural Network,VDCNN)to replace the CNN with only one convolution layer to extract high-level text features.The VDCNN draws on the ideas of VGG and ResNet to set the convolution kernel size to 3.Superimpose the network depth,and use the Shortcut and Batch Normalization methods to accelerate the convergence speed of the model.The global feature of the text is extracted by using DBLSTM(Deep Bidirectional Long Short-Term Memory,DBLSTM)with 4 layers of hidden layer,and the local features of the text extracted by VDCNN are merged with the global features extracted by DBLSTM.At the same time,the experiments were carried out on the SogouCA dataset,and compared with the single model,other deep neural network models and C_BiLSTM models,the experiments show that the proposed model further improves the accuracy of text classification.