Chinese Sentiment Analysis Based On TCN-BiGRU Model

With the continuous development of network technology and the increasingly perfect communication functions of mobile terminals,more and more network users are accustomed to expressing their opinions and suggestions on social media.These comments are very personal and reflect what the reviewer was thinking at the time.Text sentiment analysis is to dig deeply into the comments and analyze the emotional tendency expressed in them.It plays an important role in personalized recommendation services in the commercial field and public opinion monitoring in government departments.At present,there are three main research methods for sentiment analysis,which are the emotion dictionary method,the traditional machine learning method,and the deep learning method.The method of emotion dictionary has high requirements for sentiment dictionary,and the increase in text data makes it more difficult to update the sentiment dictionary.Traditional machine learning methods use training samples to learn the classifier,but the feature extraction process is complicated.Deep learning methods generally extract text statement features through CNN,RNN,LSTM,and other models and then use backpropagation to update parameters to achieve the purpose of model optimization.However,CNN ignores the sequence characteristics of the statement itself,and the problem of gradient disappearance occurs when RNN is propagated back.The excessively complex structure of LSTM leads to an increase in computation.In addition,unidirectional models cannot take context information into account and do not understand statements adequately.In view of the above problems,the following contents are studied in this dissertation.(1)In this dissertation,the TCN model and BiGRU model are fused to get the TCN-BiGRU emotion classification model.The causal convolution in TCN"memorizes" the input at the previous time and uses it for the output later,which ensures the preservation of historical information and the sequence nature of the text.In addition,with the same number of network layers,the extended convolution can obtain a larger receptive field,which can reduce the number of network parameters and prevent the over-fitting of TCN to a certain extent.GRU introduces a gate mechanism based on the RNN,which can effectively solve the problems such as gradient disappearance.At the same time,compared with LSTM,the GRU structure is more simplified and the calculation amount is reduced.BiGRU summarizes the text context information more comprehensively and has a more detailed understanding of the text.All of those make the TCN-BiGRU model have stronger learning ability.In the hotel review dataset,the TCN-BiGRU model has a higher accuracy of sentiment classification than other deep learning models,reaching 89.7%.Compared with BiGRU and BiLSTM,the accuracy is increased by 1.2%and 1.8%respectively.Compared with the machine learning model SVM,the improvement effect is also very obvious.In addition,the attention mechanism,when analyzing texts,gives different attention to each word according to its importance,which is more consistent with the characteristics of statement analysis.Therefore,this dissertation adds the attention mechanism based on the TCN-BiGRU model,and the model performance is further improved.(2)The TCN-BiGRU model is obtained by the fusion of two models.The superposition of models makes the network structure complicated,which may lead to overfitting and network degradation,affecting the performance of the model.In addition,the increase in the number of parameters in the fusion model will lead to an increase in training time.To solve these problems,this dissertation introduces the DenseNet idea into the TCN-BiGRU model and obtains the Dense-TCN-BiGRU model by referring to the DenseNet model in the computer vision field.DenseNet's dense connection mode has a certain regularization function,which further alleviates the overfitting problem.Meanwhile,feature utilization is improved,which contributes to the improvement of model performance.The introduction of a dense connection module also reduces the number of total parameters in the whole model,and the training speed and time are optimized.In the microblog comment dataset,the classification accuracy of the Dense-TCN-BiGRU model increased by 1.7%compared with the TCNBiGRU model,and the Dense-TCN-BiGRU model used fewer parameters,which accelerated the model training speed.Compared with the TCN-BiGRU model,The time of a single training session was reduced by about 1.83 seconds.The experimental results fully prove the superiority of the proposed model.Figure[21]Table[11]Reference[67]...