Abstract

This article presents an exploration of the use of pre-trained Large Language Models (LLMs) for crisis classification to address labeled data dependency issues. We present a methodology that enhances open LLMs through fine-tuning, creating zero-shot and few-shot classifiers that approach traditional supervised models in classifying crisis-related messages. A comparative study evaluates crisis classification tasks using general domain pre-trained LLMs, crisis-specific LLMs, and traditional supervised learning methods, establishing a benchmark in the field. Our task-specific fine-tuned Llama model achieved a 69% macro F1 score in classifying humanitarian information–a remarkable 26% improvement compared to the Llama baseline, even with limited training data. Moreover, it outperformed ChatGPT4 by 3% in macro F1. This improvement increased to 71% macro F1 when fine-tuning Llama with multitask data. For the binary classification of messages as related vs. not related to crises, we observed that pre-trained LLMs, such as Llama 2 and ChatGPT4, performed well without fine-tuning, achieving an 87% macro F1 score with ChatGPT4. This research expands our knowledge of how to exploit the potential of LLMs for crisis classification, representing a great opportunity for crisis scenarios that lack labeled data. The findings emphasize the potential of LLMs in crisis informatics to address cold start challenges, especially critical in the initial phases of a disaster, while also showcasing their capacity to attain high accuracy even with limited training data.