Categorizing Disaster Tweets Using Learning Based Models for Emergency Crisis Management

Abstract

Social media communication is essential to the crisis response aftermath of a massive tragedy. Facebook, Twitter, and other social media network platforms are effective instruments for connecting and fostering collaboration among catastrophe victims and other groups. As a result, numerous research publications on tweet analysis have been released. Tweet analysis during a crisis helps in understanding the nuances of the incident. Existing works primarily focused on tweet sentiment analysis and binary categorization of tweets into catastrophe relevant or not. Our work mainly categorizes catastrophe tweets into seven categories: Blizzard, earthquake, flood, hurricane, tornado, wildfire, and not-relevant tweets. Deep learning and machine learning methods were employed to categorize the tweets. The annotated data is subjected to classification using Support Vector Machine (SVM) utilizing Term Frequency-Inverse Document Frequency (TF-IDF) Vectorizer and Word2Vec Vectorizer and compares the accuracy of different kernel functions. Bidirectional Long Short Term Memory (Bi-LSTM) is used on the labeled data as a deep learning technique. SVM exhibited 88% accuracy compared to 87% for Bi-LSTM. Empirical evidence shows that our methodology is more productive and efficient than previous approaches. From this knowledge of the incident, emergency aid organizations may draw conclusions and act immediately.