| Today,e-commerce is closely related to people’s daily life.The e-commerce platform provides a comment area for users to comment online.Comment data will be stored on the e-commerce platform for a long time.For consumers,these comments directly affect whether to choose the product.For the merchants,it is possible to understand the evaluation of the product and to improve the deficiencies in time.However,the evaluation content accumulated in the comment areas on various e-commerce platforms is very large,it is very laborious to read the comments directly and obtain effective information,which not only reduces the efficiency of users reading comments,but also makes it difficult for merchants to obtain user feedback information,increasing the difficulty of improving the product.Therefore,how to efficiently obtain consumers’ emotions in such a huge amount of review data and obtain valuable review information in real time has important research significance and commercial value for improving user experience and improving product quality for merchants.Based on the BERT algorithm,this thesis obtains the evaluation subject and emotional polarity of user comments by identifying and evaluating the subject’s sentimentanalysis on the entity of the opinion and the evaluation subject,and realizes fine-grained sentiment analysis.Major tasks and contributions include:In word embedding method and collection and processing of training data,by analyzing the shortcomings of using traditional Word2Vec（Word To Vector）word vector,a word embedding method combining word vector and word vector for Chinese comments is proposed.Compared with the word embedding method using word vector and word embedding alone,the proposed word embedding method has higher classification accuracy,more complete semantic information retention,fewer unregistered words,and improved word embedding quality.By analyzing the deficiencies of ordinary datasets,the construction of high-quality labeled datasets is proposed.Comparing the experimental results of high-quality labeled datasets and common datasets,it is foundthat high-quality labeled datasets can improve the quality of model training and the prediction accuracy.In evaluating opinion fragments and evaluating subject recognition,in order to solve the problem that the traditional Word2 Vec model can only generate word vectorswith fixed semantics,a BERT-BIGRU-CRF model based on the improved concealment strategy BERT（Bidirectional Encoder Representations from Transformer）is proposed.The model improves the hidden language model of the BERT algorithm according to the close relationship between words and word groups in Chinese comments,and uses it as the language model of the word embedding layer.The word vectors trained on this model can better mine the semantic information between words and obtain more complete semantic information.In the feature extraction layer,Bi GRU is used instead of the traditional Bi LSTM to improve the training speed while ensuring the accuracy.Compared with the commonly used named entity recognition model LSTM-CRF,the proposed model has improved F1 value,recall rate and precision rate.In the evaluation of subject sentiment analysis,for the traditional RCNN model is not enough for the feature extraction of text sentiment-related information,and the classification is inaccurate,the M-BERT-ATT-Res GCNN model of word vector fusion is proposed.The model first uses the pre-trained BERT language model to replace the Word2 Vec language model,and improves it for the e-commerce review data.The improved model is M-BERT;then it uses two layers of Bi GRU to improve its ability to extract features,The residual connection is used between the two layers to prevent the gradient from disappearing;finally,the attention mechanism is used to extract the key information.The proposed model is compared with several commonly used sentiment analysis models,and the results show that the accuracy of the proposed model is improved by nearly one percentage point compared to the current better RCNN（Regions with Convolutional Neural Networks features）model. |
