Research On Sentiment Analysis Of Short Text Based On BERT And Composite Network

With the continuous development of computer science and technology and the gradual popularization of the Internet,the number of various social media platforms is gradually increasing,and the data of short texts is increasing exponentially,and the research heat of sentiment analysis of short texts is increasing.In this thesis,the text pre-training model and neural network are studied.The BERT model is used to vectorize the input text data,and the deep neural network is constructed to train and optimize the word vector.Two Chinese emotion analysis models BERT-CNN-Bi GRU and BERT-GAT are designed and implemented.Aiming at the problem that the current sentiment analysis model can not combine the local feature information of text with contextual semantic information,this thesis proposes a BERT-CNN-Bi GRU composite network sentiment analysis model.In this model,the data in the Chinese microblog data set is vectorized by BERT pre-training model,and the output word vector is trained by Text CNN and Bi GRU networks.The outputs of the two networks are then weighted and summed by attention mechanism.In this way,Text CNN’s ability to extract local feature information is combined with Bi GRU’s ability to extract context-related information.Finally,full-connection layer is used for classification.Experimental results show that the classification accuracy and F1 value of BERT-CNN-Bi GRU model proposed in this thesis are 0.9683 and 0.9681 respectively on Chinese microblog data set,which has obvious advantages over other models.Aiming at the problem that traditional neural networks cannot extract the internal structure information of text and the classification effect is reduced on the data set with complex emotion polarity,this thesis proposes a BERT-GAT Chinese emotion analysis model based on multi-level text heterogeneous graph.The model of Chinese essay in this data set the whole sentence structure split and construct "the sentence,clause," the text of the multi-level heterogeneous graph,Graph attention network training is utilized to extract a multi-level heterogeneous structure information of graph,through the selfattention mechanism will output the structure information and BERT said the sentence semantic information fusion,The enhanced text semantics represent the information and are used for final classification.Experimental results show that compared with other models,the classification accuracy of BERT-GAT model proposed in this thesis improves by 1.9 percentage points on the hotel review data set.