A White-box/Black-box Dual-modal Structured Query Method for Flood Disaster Spatiotemporal Knowledge Graph Question Answering

Abstract

Flood disaster emergency question answering (QA) requires both structured fact retrieval and complex compositional reasoning. Existing large language model (LLM)-based knowledge graph QA methods still struggle to balance query flexibility and execution reliability. To address this contradiction, this paper proposes a white-box/black-box dual-modal structured query method for flood disaster spatiotemporal knowledge graph QA. The method takes the ReAct reasoning mechanism as its scheduling core, integrating LLM-generated dynamic Cypher queries (white-box path) and predefined scripted queries (black-box path) into a unified reasoning chain. For standardized, high-frequency fact queries, the system preferentially employs black-box scripts; for tasks involving causal chain tracing, trend statistics, and integrated decision-making, the appropriate query or mixed retrieval path is invoked based on question characteristics. A query routing strategy is designed, and a data flow/control flow separation mechanism is introduced to manage large-scale query results through a "summary–file bypass" approach, alleviating context bloat during multi-round reasoning. Based on a flood disaster spatiotemporal knowledge graph and emergency response plan corpus, 30 test questions are constructed around three complex task types—causal analysis, trend assessment, and integrated decision-making—with current status query cases used to validate the standard script path. Experimental results show that the dual-modal method achieves a 100% query success rate, 98.3% answer accuracy, and an average reasoning round count of 1.47, outperforming both the pure white-box and pure black-box methods overall.