Answers are grounded in this paper and cite their sources.
To address dynamic perception lag and multi-source integration challenges in flood emergency response, this study proposes STG-GraphRAG, a closed-loop framework integrating automated long-time-series flood knowledge graph construction with geography-constrained retrieval-augmented generation. Using flood-related social media, images, and meteorological data from July 2024, we construct 31 daily snapshots of a heterogeneous flood knowledge graph. For each query, keyword matching and semantic embeddings identify candidate nodes, which are then filtered by spatial and temporal constraints. Geography-aware community expansion uses spatial proximity and attribute similarity to form an evidence subgraph for answer generation. Comparison of location mentions in social-media records and official documents showed different coverage across administrative levels; Hekou Town appeared in the social-media records on July 29 and in the compared official documents on July 31. This method enhances context relevance to 0.9818 and compresses the subgraph scale by 68.5%. By integrating spatio-temporal summarization and risk assessment modules, the system derives an internal risk-priority category from hotspot, emotion, and flood-event indicators. In the constructed benchmark, the system achieves an answer relevance of 0.7705, a faithfulness of 0.9687, and a human score of 4.4146 (out of 5). On the pooled FloodBot benchmark, STG-GraphRAG showed higher observed answer-relevance and human-score means than INITIAL. The link of this project: https://github.com/muhan-yy/STG-GraphRAG. CONTACT Xiaoping Rui ∗ ruixpsz@163.com
No takes yet. Share an insight, caveat, or question.
Tang et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: