2026-11-06 テキサスA&M大学
Texas A&M researchers are developing an AI-based tool to help emergency managers quickly sort through the flood of information that emerges during crises such as hurricanes. Credit: Courtney Sacco/Texas A&M AgriLife
＜関連情報＞
- https://stories.tamu.edu/news/2026/11/06/researchers-develop-ai-tool-to-improve-crisis-communication/
- https://www.sciencedirect.com/science/article/abs/pii/S1474034626005288
CrisisSense-LLM：災害情報学におけるマルチラベルソーシャルメディアテキスト分類のための、指示に基づいて微調整された大規模言語モデル CrisisSense-LLM: instruction fine-tuned large language model for multi-label social media text classification in disaster informatics
Kai Yin,Bo Li,Chengkai Liu, Ali Mostafavi,Xia Hu
Advanced Engineering Informatics Available online: 28 May 2026
DOI:https://doi.org/10.1016/j.aei.2026.104836
Abstract
In the field of disaster informatics, social media is increasingly used for improving situational awareness to inform response and relief efforts. Efficient and accurate text classification has become a central area of investigation. This study proposes a novel approach to disaster text classification by instruction fine-tuning a pre-trained Large Language Model (LLM) for unified multi-label classification of disaster-related tweets. This study constructs a comprehensive instruction dataset with more than 60,000 disaster-related tweets and uses it to fine-tune an open-source LLM with disaster-specific knowledge. The resulting model simultaneously classifies event type, informativeness, and humanitarian aid type, enhancing the utility of social media data for disaster situational awareness. The fine-tuned model achieves an overall exact-match accuracy of 63.8%, outperforming a supervised BERT baseline (57.67%). This improvement is consistent across all three classification tasks and is especially notable for humanitarian aid type classification, the most challenging of the three tasks, where the model achieves substantial gains over the supervised baseline, demonstrating the superior capability of the fine-tuned LLM model to distinguish semantically nuanced humanitarian categories in disaster-related social media posts. This research offers a useful step towards integrating LLMs into disaster management tools for improved situational awareness and response support.
