Disaster Question Answering with LoRA Efficiency and Accurate End Position

Researcher Takato Yasuno has introduced a novel disaster-focused question-answering system designed to address the critical need for accurate, domain-specific information during natural disasters like earthquakes and floods. The system tackles a core problem: when people face rare, high-stress disaster situations, they lack experience and reliable knowledge. Standard RAG (retrieval-augmented generation) with large language models can fail to find relevant information and may hallucinate, potentially spreading dangerous misinformation. This new model is specifically tailored to Japanese disaster contexts and response experiences.

The technical architecture is built on cl-tohoku/bert-base-japanese-v3, enhanced with a Bi-LSTM layer for better contextual understanding and 'Enhanced Position Heads' for precise answer span detection. Its standout feature is efficiency, achieved through LoRA (Low-Rank Adaptation) fine-tuning. The model achieved a 70.4% End Position accuracy and a strong 0.885 Span F1 score while utilizing only 6.7 million parameters—just 5.7% of the base model's 117 million. This makes it suitable for real-world deployment where computational resources may be limited. Future work outlined includes creating benchmark datasets, further fine-tuning foundation models with disaster knowledge, and developing lightweight, power-efficient edge AI applications for use when communication and power are scarce during actual disasters.