With the rapid development of the Internet, Internet public opinion has received more and more attention. Natural disasters are sudden and real-time. However, the current stage of Internet public opinion is often divided only by the size of the public opinion, which can't accurately applied to public opinion on natural disaster networks. This paper uses the similarity calculation method to consider the impact of text content semantics on the division of Internet public opinion stage, and proposes a scheme for adaptively dividing the time series based on text similarity. The experiment finally completed the dynamic time series division of Internet public opinion, followed the public opinion hotspots in real time during the development of the incident, studied the evolution of Internet public opinion hotspots in different stages, and provided a reference for the government to monitor the changes in the Internet public opinion hotspots in real-time and formulate a response decision.