Knowledge Graph-Driven Training Data Construction for Urban Flood-Traffic Scenario Generation Using Small Language Models

Urban flooding caused by extreme rainfall events disrupts transportation systems, yet generating realistic flood-traffic scenarios for disaster preparedness remains a labor-intensive manual process. This study proposes a Knowledge Graph (KG)-driven pipeline that automatically generates domain-specific training data for fine-tuning small language models (sLLMs) to synthesize urban flood-traffic scenarios. A domain KG comprising 58 entities and 285 relationships was constructed for Jinju City, South Korea, integrating empirical flood data from 112 local documents with quantitative rainfall-traffic impact values from 14 international studies. Nine domain constraint rules, including a novel spatial consistency rule, ensure the physical plausibility of generated scenarios. Through constrained weighted graph walks, 800 semi-structured English narrative scenarios were automatically generated in approximately 5 min, substantially reducing the labor required compared to manual creation. Three sLLMs spanning different architectures and parameter scales—Flan-T5-Large (770M), Qwen2.5-3B-Instruct (3B), and Qwen2.5-7B-Instruct (7B)—were fine-tuned using QLoRA on a single GPU with 16 GB VRAM. Evaluation on 78 test samples demonstrated consistent performance improvements with increasing model scale: Qwen2.5-7B achieved BLEU-4 of 0.5524, ROUGE-L of 0.6883, BERTScore F1 of 0.9662, and KG Fact Consistency of 1.0000, representing a 33.8% BLEU-4 improvement over Flan-T5-Large. Both Qwen models achieved KG Fact Consistency of 1.0000. The 3B model achieved 98.6% of the 7B model’s BLEU-4 at 53% of the VRAM cost with identical factual consistency, representing the most cost-effective configuration. All models were trained for 10 epochs on the same GPU, demonstrating practical feasibility for municipal disaster response deployment.