Multi-Task Learning Model for Fine-Grained Categorization of Emergency Tweets

Information processing in emergencies needs to be fast and accurate. Social media platforms, such as Twitter, have become an important source of real-time information in times of crises. However, the sheer volume of tweets necessitates automated methods for effective categorization. This paper introduces a multi-task learning (MTL) approach for classifying emergency-related tweets. We will train our model to predict both the type of emergency and the urgency level of a tweet simultaneously, exploiting the inherent correlation between the two to potentially increase general performance. We report experiments of our approach on a collection of crawled tweets from various emergencies that we procured primarily from CrisisNLP benchmark datasets. The findings indicate that multi-task learning outperforms single-task learning approaches and obtains higher accuracy and F1-scores over both tasks at hand. We also do fine-grained error analysis to draw attention to the challenging cases and propose directions for future work: integrating external knowledge, handling multilingualism and code-switching, data imbalance, and model explainability. Our work has shown the potential of MTL in improving emergency response through enhanced social media analysis and lays down a solid foundation for future real-time implementations of such systems.