This paper presents a novel approach that overcomes the limitations of existing single-label classification models by considering the multifaceted and dynamic nature of disaster-related social media data. We introduce a directive fine-tuning technique that utilizes a large-scale language model (LLM) to perform multi-label classification of disaster-related tweets. We build a comprehensive directive dataset from disaster-related tweets and use it to fine-tune an open-source LLM to incorporate disaster-specific knowledge. The fine-tuned model simultaneously classifies information across multiple dimensions, including disaster type, informational value, and humanitarian intervention, significantly enhancing the utility of social media data for disaster situational awareness. Experimental results demonstrate that this approach improves the classification of critical information in social media posts, making it more effective for situational awareness in emergency situations. This contributes to the development of more advanced, adaptive, and robust disaster management tools that enhance real-time situational awareness and response strategies.
