Leveraging advanced deep learning models for disaster response

Abstract

At the beginning of natural disasters, obtaining timely and accurate information is of great importance for effective response and recovery efforts. Social media platforms like Twitter play a critical role in providing real-time updates, eyewitness accounts, and requests for assistance. Rapidly assessing this information, including identifying informative tweets and their categories, is vital for directing resources and prioritizing rescue operations. Deep learning (DL) methods have proven effective in disaster classification tasks involving both text and image data. Building upon this foundation, my research aims to further leverage advanced deep learning models to enhance the classification of disaster-related posts on X (formerly known as Twitter). Given the challenges associated with labeling datasets in the immediate aftermath of a disaster—when data needs to be analyzed in real-time for urgent damage and situational awareness information, and manual annotation by experts is time-consuming—I have investigated methods suitable for handling smaller labeled datasets. These methods were evaluated in both in-domain and cross-domain settings to assess their effectiveness across datasets with similar and different characteristics. Additionally, the performance of the models was explored in few-shot and zero- shot settings, acknowledging the scarcity of labeled data as a disaster unfolds. In particular, the potential of Large Language Models (LLMs), due to their generalization capabilities from extensive pretraining and scale, was explored for zero-shot and few-shot settings. Results showed that LLMs are invaluable for quick response efforts in disaster situations at the beginning of a crisis. These advancements aim to improve situational awareness and optimize resource allocation during disaster events, ultimately contributing to the reduction of human and economic tolls.