2026 年 7 巻 2 号 p. 20-41
This study presents a pipeline for automatically generating natural-language descriptions of disaster situations from GIS spatial data using large language models (LLMs). The pipeline converts GeoJSON landslide distribution data and location reference information into structured JSON through spatial aggregation, which is then provided to an LLM to produce per-district damage descriptions. A verification module detects numerical errors and unsupported content in the generated text. Through a controlled comparison of prompt conditions, this study demonstrates that prohibition instructions—explicit directives forbidding the model from introducing information not present in the input data—reduce unsupported content and improve faithfulness (Noto: 3.60 to 4.64; Iburi: 3.84 to 4.76), without degrading fluency or style appropriateness. However, prohibition instructions also suppress qualitative interpretations of numerical data (Interpretation Adequacy: Noto 3.54 vs. 4.11; Iburi 3.54 vs. 4.20), revealing a trade-offbetween hallucination control and interpretive richness. A model scale comparison shows that 27–32B-parameter open models (4-bit quantization, locally executable) achieve practical quality with faithfulness above 4.6, whereas 7B-class models remain clearly below the 27–32B models and fall below 4.0 on the Noto dataset. The pipeline was applied without modification to the 2024 Noto Peninsula Earthquake (430 districts) and the 2018 Hokkaido Eastern Iburi Earthquake (51 districts), yielding comparable performance across these two landslide case studies. An expert evaluation by three disaster-prevention practitioners reproduced this trade-offand additionally showed that actionability and information hierarchy—two aspects not captured by the automated metrics—were rated higher without prohibition, indicating that the faithfulness gained from prohibition carries a measurable cost in operational usefulness.
