[U-11] Compound Disasters across Atmosphere, Hydrosphere and Geosphere

[U11-03] Understanding of cascading disaster and countermeasures in the event of complex and catastrophe disasters

★Invited Papers

*HIROI U¹ (1.The University of Tokyo)

Keywords:cascading disaster, machine learning

In the event of a complex and catastrophe disaster, disaster chains are known to occur when one event causes another one to occur repeatedly, eventually affecting completely different areas as well. Studying disaster chains is an effective means of identifying possible future disaster events in advance or in real time. To reach this goal, the authors of the present study developed in a mechanical method to create a cascading disaster networks from Japanese newspaper articles. In order to extract causal sentences a machine learning discriminant model for candidate causal sentences extracted using both cue phrases and succession expressions with causation was created. For machine learning, a support vector machine is used to make the computer learn the combination of a sentence annotated with the presence or absence of a causal relationship and the syntactic and semantic features of the sentence, and then the presence or absence of a causal relationship is estimated from the syntactic and semantic features of the sentence for which the presence or absence of a causal relationship is unknown. In this study, it was found that causal sentences could be extracted from disaster articles with a certain degree of accuracy. It was also possible to create a disaster causal network with sentences as nodes and links. In the cascading disaster networks derived from newspaper articles that appeared during the six months after the Great Hanshin-Awaji Earthquake and the Great East Japan Earthquake by using this method it was possible to extract knowledge of events and causal relationships that could not be obtained with the conventional cascading disaster networks.