Emergencies inflict heavy casualties, economic losses, ecological damage, and significant social harm to society. By segmenting information topics and analysing emotional shifts, we can identify corresponding real-world events and their impacts, thereby providing guidance for timely responses to emergencies. In the past, public opinion monitoring of emergencies was based mainly on single-topic detection or emotion analysis, which cannot comprehensively evaluate the evolution of public opinion. In this work, word segmentation is applied to video comments related to various emergency situations. By utilizing the co-word network and Louvain algorithm for theme division, along with sentiment analysis constructed through time series analysis of sentiment value changes for various emergencies employing the naive Bayes method, the evolution of public opinion is comprehensively assessed. As a result, the pivotal nodes in the evolution of public opinion are identified and the evolution process is divided into stages. Using this method, relevant management departments can effectively address the majority of public opinions for various types of emergencies, addressing them from the perspectives of prevention, adjustment, and recovery. This approach not only enhances rescue efficiency and strengthens safety management but also actively guides the evolution of public opinion, ultimately providing society with solid and reliable security safeguards.