Abstract: Automatic classification of speech acts can help to understand the intentions and behaviors of social media users' discourse and effectively uncover public opinion. Based on the Speech Act Theory, we provide a detailed classification of social media users' intentions expressed in tweets. We manually annotate four thousand vaccine-related tweets and extract user features, temporal features, text vector features, topic features, sentiment features, etc. Then, machine learning methods such as logistic regression, random forest, XGBoost, and a combination of BERT and neural network models are used to construct speech act classification models in the context of public emergencies. The SHAP interpretation method is used to rank the importance of features. Finally, we use the Kruskal-Wallis test to evaluate the differences in sentiment and the impact of various speech acts. The accuracy of speech act classification based on the XGBoost model reaches 0.792, which is better than baseline models. Text vector features have the highest importance in speech act classification. There is no significant difference in the number of retweets among different tweet speech acts, while there are significant differences in the number of likes and sentiment features.

Key words: Speech act, Topic analysis, Public emergencies, XGboost, Sentiment analysis