Rule-Derived Address Knowledge Graph for Indoor–Outdoor Spatial Question Answering

Spatial question answering requires modeling relations among entities, yet metric georeferencing faces structural constraints in routinely available operational data: two-dimensional coordinates cannot distinguish floors within a building, rule-defined address relations require geometric approximation after coordinate conversion, and unregistered indoor spaces lack official geometry. We propose question answering over a knowledge graph derived deterministically from the assignment rules of the Korean road-name address system, in which number parity, order, and difference encode the opposite-side, same-side successor, and address-order interval relations. Mapping these rules onto graph edges yields a three-layer graph—reference frame, physical entities, and facilities—that resolves relational queries by edge traversal without coordinate computation. On a graph of 191,808 entities and 380,743 relation edges for Jongno-gu, Seoul, 93.9% of the rule-derived opposite edges were geometrically confirmed, and on an independently constructed 400-question benchmark the proposed method reached an F1 score of 0.381—under a structural ceiling imposed by administrative–physical binding coverage—versus 0.018 for relation-ablated retrieval and 0.103–0.119 for relation-verbalized retrieval. Queries on unregistered floors, which lack official geometry, reached an F1 score of 0.647; the advantage persisted through answer generation with three open 7–8B models (0.316–0.359 vs. 0.004–0.017); and applications to Sydney, Seattle, Paris, and Amsterdam yielded 0.521–0.746 across address-assignment schemes.