Disaster Question Answering with LoRA Efficiency and Accurate End Position

TL;DR

This work introduces a disaster-focused question answering system based on Japanese disaster situations and response experiences and demonstrates that the combination of Japanese BERT-base optimization and Bi-LSTM contextual understanding achieves accuracy levels suitable for real disaster response scenarios.

Abstract

Natural disasters such as earthquakes, torrential rainfall, floods, and volcanic eruptions occur with extremely low frequency and affect limited geographic areas. When individuals face disaster situations, they often experience confusion and lack the domain-specific knowledge and experience necessary to determine appropriate responses and actions. While disaster information is continuously updated, even when utilizing RAG search and large language models for inquiries, obtaining relevant domain knowledge about natural disasters and experiences similar to one's specific situation is not guaranteed. When hallucinations are included in disaster question answering, artificial misinformation may spread and exacerbate confusion. This work introduces a disaster-focused question answering system based on Japanese disaster situations and response experiences. Utilizing the cl-tohoku/bert-base-japanese-v3 + Bi-LSTM + Enhanced Position Heads architecture with LoRA efficiency optimization, we achieved 70.4\% End Position accuracy with only 5.7\% of the total parameters (6.7M/117M). Experimental results demonstrate that the combination of Japanese BERT-base optimization and Bi-LSTM contextual understanding achieves accuracy levels suitable for real disaster response scenarios, attaining a 0.885 Span F1 score. Future challenges include: establishing natural disaster Q\&A benchmark datasets, fine-tuning foundation models with disaster knowledge, developing lightweight and power-efficient edge AI Disaster Q\&A applications for situations with insufficient power and communication during disasters, and addressing disaster knowledge base updates and continual learning capabilities.

Figures (5)

• Figure 1: Overview of Disaster-specific Q & A System. The diagram illustrates the comprehensive architecture of our disaster question answering system, showing the integration of Japanese BERT-base processing, Bi-LSTM contextual understanding, enhanced position prediction heads, and LoRA optimization components. The system processes disaster-related queries through specialized preprocessing, contextual encoding, and precise answer extraction optimized for emergency response scenarios.
• Figure 2: Performance Comparison: Our BERT-base + Bi-LSTM model achieves 70.4% End Position accuracy, substantially outperforming baseline models including BERT-base-only (45.3%), lightweight variants (15.0%), and various training sample configurations (6.0%-9.5%).
• Figure 3: Performance Trend Analysis: Evolution of model performance across development phases, showing consistent improvements from baseline implementations to our final BERT-base + Bi-LSTM architecture with LoRA optimization.
• Figure 4: Algorithm Flow Part 1: Multi-modal Input Processing and Context Encoding. The diagram shows the initial stages of our system including question preprocessing, disaster context analysis, Japanese BERT-base tokenization, and contextual embedding generation. Key components include specialized tokenization for disaster terminology, contextual relationship mapping, and attention-based feature extraction optimized for emergency response scenarios.
• Figure 5: Algorithm Flow Part 2: Enhanced Position Prediction with LoRA Optimization. The diagram illustrates the advanced processing stages including Bi-LSTM sequential modeling, Enhanced Position Heads for precise boundary detection, and LoRA-based parameter-efficient optimization. The workflow shows how contextual embeddings are processed through bidirectional LSTM layers, enhanced position prediction mechanisms, and final answer span extraction optimized for disaster response accuracy.