An English-Japanese Twitter-Based Analysis of Disaster Sentiment during Typhoons and Earthquakes

Having proper situational awareness during disaster situations is crucial in planning and mitigation. Knowing people's perception, needs and behavior during disasters is critical in developing the right management strategies. However, cities with multilingual and diverse international populations may react differently to disasters and gaps still exist in understanding this issue. Microblogging with social media has become a prevalent tool during emergencies and disasters. In this paper, we present a method in analyzing the sentiment of both the local residents and foreigners in Tokyo during case studies of earthquake and typhoon. Through the use of Twitter data, we retrieve individual tweets specifically on the onset of the disaster both in Japanese and English. After preprocessing, we develop Machine Learning algorithms using Support Vector Machine and XGBoost to classify tweet sentiment. The sentiment analysis models obtain fair accuracy and could be scaled and applied to sentiment classification of tweets during other types of natural disasters. Moreover, since our model is trained specifically on disaster tweets, it could yield a more accurate and contextual result when applied to future disasters. Furthermore, we analyzed information through Word Cloud, keyword analysis and time series analysis of sentiment polarity. We deduce that Japanese show a more positive sentiment than foreigners at times of disaster. Additionally, we observe that negative sentiment of both groups is lower for typhoons than earthquakes overall. Lastly, using this methodology could provide insights specific to typhoon and earthquake contexts to elicit requirements for disaster information or warning systems catered towards foreigners in the area which could be used in disaster management.