[Purpose/significance] When the public is threatened by sudden public events, it is easy to create a crisis of trust in government agencies and exacerbate public opinion risks. Emotions are a potential driving force behind the development of public opinion, but there has been little research on the public's emotional risk warning, and there is no detailed definition or judgment. [Method/process] Based on the theory of risk communication, this article evaluates the degree of public emotional risk warning from a qualitative perspective, designs an emotional risk feature system from a quantitative perspective, constructs an emotional risk identification and warning framework, and analyzes the warning performance of the model and the degree of contribution of different features from multiple angles. [Result/con? clusion] The experimental results found that Logistic Regression performed well in emotional risk identification tasks; the combination of TextRCNN and BiLSTM + Attention had the best performance and could efficiently complete the identification and warning tasks for emotional risk; feature ablation experiments found that both emotional and seman? tic features had a significant impact on model performance, while influence features had a low contribution rate; text vector features could improve the risk perception performance of deep learning models.