Identifying Panic Triggers from Disaster-Related Tweets

Using social media platforms such as Twitter has drastically increased over the past decade. It has enhanced the traditional means of communication in many aspects of life. Artificial Intelligence and Machine Learning algorithms have become popular in assessing natural disasters. During natural catastrophic events and emergencies, people progressively use microblogging platforms such as Twitter, creating a high volume of posts spread across these platforms. The information disseminated on Twitter contains critical indicators about evacuations or emergency actions that could incite panic, affecting the response and evacuation behavior of the general population. In order to avoid panic, these indicators need to be detected, the credibility of their source needs to be validated, and the emergency agencies need to mitigate the risk of panic by quickly taking the right actions for these panic triggering situations. This paper presents a Panic Trigger Identification Method (PTIM) which applies machine learning techniques on disaster-related tweets to detect panic triggers, and classifies the tweets based on the triggers identified and the corresponding credibility level of the tweets to improve the emergency response, and to suggest mitigation actions for emergency management. Two types of text vectorizers, CountVectorizer and TfidfVectorizer, are used as features for the supervised machine learning classification models. A performance comparison is conducted among the classifiers. Results show that for the classification of the tweets with panic triggers, Random Forest and Decision Tree give the best predictions with high accuracy (95% on average) when using CountVectorizer features.