A Study of Bert-Based Sentiment Analysis Algorithm for Short Text Movie Reviews

In order to complete the analysis of the emotional value orientation problem of the user's viewing experience of the movie evaluation, it is preferred to carry out data preprocessing on the collected text and use the back translation enhancement technique in the data enhancement technique to complete the oversampling process of the negative samples, and utilize the upgraded version of the pre-training model Bert's DeBERTa model to complete the extraction of the features of the initial preconditioning, and then carry out with the vectors generated from the M3E model to Fusion, and then the obtained text feature representation is input into the Bi-LSTM network to complete the further extraction of text features, and Attention attention mechanism is added behind the Bi-LSTM network, so that the extraction of model features focuses on the more important information. Finally, the final classification task is accomplished through the classifier by outputting the text features after the initial text has passed through the M3E-DeBERTa+Bi-LSTM+Attention model. The completed models are used as BiLSTM,BERT, DeBERTa, RoBERTa, BERT-BiLSTM, RoBERTa-BiLSTM, DeBERTa-BiLSTM and DeBERTa-BiLSTM-Attention models as a comparative model with the proposed M3E-DeBERTa- BiLSTM-AT model, and the experimental results show that the accuracy of this model on the test set is improved by 21.51%, 13.49%, 11.32%, 8.06%, 7.27%, 4.88%,2.11%, and 1.3%, respectively, which indicates that the M3E-DeBERTa+Bi-LSTM+Attention model can be effective in improving the accuracy of the problem about user's sentiment classification for movie evaluation.