DeLTran15: A Deep Lightweight Transformer-Based Framework for Multiclass Classification of Disaster Posts on X

Abstract

During disasters, timely and accurate information is paramount for effective decision-making and resource allocation. Social media (SM) platforms, particularly X platform (formerly Twitter), have become crucial sources of real-time data during catastrophic events, offering insights into evolving situations, needs, and responses. However, efficiently detecting and extracting meaningful content from the massive amounts of unstructured user-generated data swiftly is challenging. While large transformer models like BERT perform exceptionally well across a range of natural language processing tasks, including SM text processing, their computational demands make them impractical for real-time, low-resource applications. Therefore, developing lightweight yet effective models to detect and categorize relevant SM information during catastrophic events is essential for rapid response efforts. To address this need, a novel lightweight framework 'DeLTran15' is introduced, designed specifically for classifying posts on X shared during disasters into multiple humanitarian information categories. The framework uses the 'Obtain Scrub Explore Model iNterpret' (OSEMN) methodology, which enhances the reliability and effectiveness of the classification process, leading to a robust and high-quality final model. Six lightweight pretrained language models: BERT-tiny, BERT-mini, ALBERT-base, XtremeDistil, Electra small discriminator (ESD), and Electra small generator (ESG), all under 15 million parameters are finetuned. Extensive experimentation on a benchmark real-world disaster dataset indicates XtremeDistil's superior performance among all lightweight models. It is further optimized using post-training quantization techniques, enabling deployment on resource-constrained devices. Contrary to the state-of-the-art larger transformer language models with over 100 million parameters, DeLTran15 achieves high classification accuracy with enhanced computational efficiency, making it suitable for real-time disaster management.