Social media contributes to enhancing transparency and openness for the purpose of innovating public services and policy-making. In disaster management, social media data can be mined to discover public perceptions and concerns on large disaster events. However, converting large data streams into useful information remains a challenge due to the unstructured nature of textual data. ?is study proposes an interactive topic modeling method to analyze microblog data for understanding the dynamics of public expressions immediately a?er a major explosion event. First, we extract topics from microblog message data. In order to test the influence of the number of topics, the topics are detected at multiple levels of granularity by varying the number of topics. Second, these topics are used to detect topical compositions of contents at different time slices and assess the topic evolution over time. The topic evolution patterns are visualized by the streamgraph method to discover informative topics to help to take further actions. Third, since the first-level topics are not informative, we conduct a second-level topic (subtopic) analysis to detect key decision elements by choosing "investigation" from the first-level topics, a hot focus in any man-made disaster. The results improve our understanding of the topic composition evolution around major man-made disasters and have implications on officials deciding what and when to release formal investigation information to the public.