BERT-based Sentiment Analysis of Chinese Online Social Movements

Online social movements are a group of netizens with the same or similar purpose, spontaneously discussing and disseminating certain information, trying to attract more people to participate, and creating a public opinion on the Internet or even a social atmosphere of anxiety. Online social movements analysis can reveal the evolution of sentiment during the movements and therefore prevent the relevant social disasters from happening. The literature shows the lack of public sentiment analysis text datasets and effective analysis methods for Chinese online social movements. In this paper, we classify sentiment into four categories: positive, anger, anxiety, and weak negative. We believe anger and anxiety are the two most important sentiments in forming an online social movement. Afterwards, we first time create a public sentiment analysis dataset about Chinese online social movements, and then propose a bidirectional encoder representation from transformers (BERT)-based model to classify the sentiment. Moreover, we use the focal loss in the BERT model rather than cross entropy loss to enhance the contribution of minority classes to the total loss to address the imbalance issue in the dataset. The proposed BERT-based model is compared with six baseline methods. The results show that our method outperforms those baseline models by achieving higher macroF1 scores.