CURD: Context-aware Relevance and Urgency Determination

Ademola Adesokan, Sanjay Madria

Abstract: During emergencies where time is of the essence, efficient management of disasters depends on swiftly recognizing relevant and urgent information from online platforms like X (Twitter), which is imperative for augmenting established response frameworks, such as the 911 emergency system. This paper introduces CURD, a Context-aware Relevance and Urgency Determination system designed to enhance the efficiency of disaster response. The system addresses two critical challenges: filtering out irrelevant data and assessing the urgency of relevant information. Our approach includes a multi-level annotation process for event type, relevancy, and an urgency annotation algorithm that significantly improves information extraction accuracy and efficiency. CURDdl, our classifier, uses a deep learning pipeline architecture with a combination of transformer models, a convolution layer, and custom attention mechanisms to classify disaster-related tweets into multiclass-event type, binary-relevance-and-urgency categories, and rank urgent ones based on significance. Experimental results show that our best baseline classifiers for all three tasks achieved ≥ 88% F1 and accuracy, and ≥ 94%. AUC. Our models also outperformed models from related works in all metrics, validating the effectiveness of CURD in prioritizing response messages that will facilitate decision-making and resource allocation in disaster scenarios. CURD annotated dataset and code are available on GitHub 1.