Deep learning based sentiment analysis during public health emergency
Zezhong Pan, Wenchao Xu · 2021
The traditional method of text sentiment analysis mainly uses sentiment dictionary or machine learning, but sentiment dictionary method needs to update the thesaurus in time for microblog, which is noisy, has many new words, uses abbreviations to express emotions and the labor cost is high. Traditional machine learning methods can not extract salient features, resulting in low classification accuracy. In order to solve these problems, this paper proposes a model BERT_RCNN for sentiment analysis of netizens during public health emergency based on deep learning. BERT_RCNN first uses BERT to fine tune the input text, and trains the text into vectors to represent it. Then it takes the trained vectors as the input features of the upstream model and learns the microblog text features through RCNN network. Experiments show that: The accuracy of BERT_RCNN is 91.80%, which is better than 72.74% of SVM, 72.03% of logistic regression and 68.95% of naive Bayesian model. Better than other deep learning models BERT_Bi-LSTM, BERT_FC, BERT_DPCNN, ERNIE_FC with higher accuracy. Because it has faster convergence speed and shorter training time. Therefore, the proposed method achieves higher recognition accuracy with lower computational cost.