Using the Bert model and the attention mechanism to obtain an accurate sentiment analysis model

This study takes the network reviews captured by Ctrip as the object of investigation. 29,025 reviews are preprocessed first, and the training set, verification set and test set are divided according to proportion. The Bert model is used for training and data parameters are retained. Only the subject and affective adverbs were retained and the negative adverbs were artificially weighted. The data set was retrained. The research results supplemented the innovation and application value of the training results, and could prove whether there were commercial malicious competition and other phenomena in bad reviews in network evaluations.