Research On Text Emotion Analysis Based On BiTCN And Pre-training

Text sentiment analysis is one of the basic tasks in the field of natural language processing.It is of great significance for consumers and businesses to research and analyze the emotional views contained in a large number of text data on the network platform through relevant algorithms.However,in the traditional emotion analysis methods,CNN can only better mine the local features of the text,and will not pay too much attention to the context information.RNNs are prone to gradient problems during training,and at the same time,the training time will be too long due to the lack of support for parallel computing.Therefore,this dissertation focuses on the TCN network which has the ability to deal with sequence problems on the basis of CNN,and further realizes text emotion analysis based on the improved model.The following is the research content of this dissertation:(1)An emotion analysis method combining bidirectional time convolution neural network and attention mechanism is proposed to solve the following problems: The mainstream emotion analysis methods cannot fully capture text features,and the analysis ability of long text is relatively weak;The expression of comment text on the network platform is colloquial,which makes the importance of words in the text indistinguishable.Firstly,use Word2 vec model to train and get word vector expression of all words in long text;Then,the bidirectional TCN is used to encode and learn the context information from two directions to obtain more sufficient text representation,in which the receptive field is changed by means of expansion convolution,and the data leakage is prevented by means of causal convolution;Then the attention mechanism is introduced to add weight information to each feature vector in the text to highlight its internal relevance;Finally,softmax is used to get the output results of the model and complete the emotion ternary classification in the dataset.The experimental results show that the accuracy of the Bi TCN-Attention model in classification can reach 85.1%,which is improved to varying degrees compared with CNN,RCNN,LSTM,TCN,Bi LSTM and Bi TCN,proving the effectiveness of this method.(2)Based on the above model,a text sentiment analysis method combined with pre-training is proposed to solve the problem that the traditional word vector method cannot fully express rich semantic information due to the long distance between the target word and the sentiment word in the text.Specific innovation: A feature extraction module(CW)integrating word vectors is given.On the one hand,this module obtains word vector feature by using Word2 vec training corpus,and on the other hand,it further extracts word vector feature by using ALBERT.Then the combined features extracted from the two aspects are used as the input of the Bi TCN-Attention layer,and the output results of the model are obtained through Softmax.The experimental results show that the classification accuracy of the CW(ALBERT)-Bi TCN model integrated with the CW module is 83.6%,which is improved compared with the model using only word vectors or word vectors as input.The classification accuracy of the CW(ALBERT)-Bi TCN-Attention model is 86.4%,which is improved compared with the model that only adds CW modules or Attention mechanism,which proves the effectiveness of this method and further demonstrates the applicability of the Bi TCN-Attention model mentioned above.Figure [22] Table [9] Reference [76]...