Self-Inhibition Residual Convolutional Networks For Chinese Sentence Classification

Inspired by the big success of deep learning methods in Computer Vision,convolutional network has become a dominant approach in many Natural Language Processing(NLP)tasks.Text classification,a very typical research area in NLP,is the foundation of many upper applications.However,these networks for classification are pretty shallow and simple compared to the advanced and sophisticated networks such as ResNet and DenseNet,so they are not able to capture the hierarchical feature of text.In addition,there are different methods for converting texts into image-like representation.The traditional text representation is commonly used bag of words,which considers whether the words appear in texts.Word2 vec is an unsupervised learning to obtain the vector representation of words,which introduce the word co-occurrence information.Text preprocessing of those models in Chinese are quite rough,which leads to the loss of rich semantic information.In this thesis,we explore deep convolutional networks for Chinese sentence classification and present a new model named Self-Inhibition Residual Convolutional Network(SIRCNN).This model employs extra Chinese character information and replaces convolutional block with self-inhibiting residual convolutional block to improve performance of deep network.At the same time,for the difficulty of network model training caused by deeper network structure,we innovatively use self-restraining residual convolution module to replace the previous convolution module to improve the comprehensive performance of the network.Finally,the fitting superiority and training superiority of the Self-Inhibition residual module are analyzed,and a comprehensive comparative experiment is carried out.Experiments show that our network model can achieve the leading effect on three of the five different fields of classification data sets,and the accuracy on Chat data sets is improved from 86.92% to 90.89%.Although the effect on two large datasets is slightly inferior to that on Bi-LSTM(Bidirectional Long Short-Term Memory),there is only a small difference of 0.3% to 0.5%.However,in the time cost of the whole training process,BiLSTM needs four times as long as SIRCNN to achieve a competitive result.Therefore,SIRCNN is still the best choice.Experiments show that the SIRCNN outperforms other network structures including traditional convolution network and residual network in terms of convergence speed and final accuracy.The whole network training process does not introduce any additional parameters,which avoids the increase of computational complexity of the model.