LAPUC-DR: LLM-AUGMENTED PROMPT-BASED URGENCY CLASSIFICATION FOR DISASTER RESPONSE

ABSTRACT

Appropriate urgency classification plays a very important role in successful disaster management. Social Media systems (Twitter, in particular) have emerged as an excellent source of real-time information; however, conventional systems lack contextual comprehension and require manual labeling, which is labor-intensive. This paper addresses the issues and proposes the LAPUC-DR framework, combining large language models with a variety of contextual enrichment methods. One of the main contributions is a zero-shot prompt-augmented annotation scheme that uses metadata, namely, event type, subevent type, topic category, and emotional tone to augment the raw text and create annotated data. Then, automatic urgency labels with high quality are created with these enriched prompts. The transformer-CNN attention classifier is tuned using the annotated data to improve the performance. Experiments conducted on a real-world disaster dataset demonstrate that LAPUC-DR achieves an F1 score of 94%, significantly outperforming existing models in all aspects across a wide range of situations, thereby exhibiting robust generalizability to unseen events. Findings ensure that the composition of multimodal context and generative-supervised pipelines yields a considerable increase in the quality of urgency classification. In conclusion, LAPUC-DR provides a real-time, scalable, and interpretable answer to disaster informatics. The usage of multimodal inputs, the most prominent of which are the images and geolocation data, will be explored to increase the applicability of the emergency response systems further on and, at the same time, to make them more efficient.

Keywords: Disaster Response, Urgency Classification, LLMs, Zero-Shot Learning, Prompt Engineering, Social Media Analysis