GNLM: A Graph-Native Language Model With Road Name Address-Based Spatial Reasoning for Geographic Question Answering

Abstract

Geographic question answering (QA) requires factually accurate responses grounded in structured spatial knowledge. However, existing approaches based on large language models (LLMs) suffer from factual hallucination, while retrieval-augmented generation (RAG) lacks structural reasoning capabilities. We propose GNLM (Graph-Native Language Model), a novel architecture that restructures the traditional LLM pipeline by performing all reasoning through knowledge graph traversal and restricting the role of LLMs to natural language generation. The knowledge graph is constructed exclusively from government public data sources (building registry, road name addresses) and expert-curated domain knowledge, ensuring factual reliability without LLM-generated content. GNLM introduces four key contributions: 1) a Graph-First, LLM-Last architecture with fact validation that structurally prevents fabrication (all observed errors stem from knowledge gaps, not false generation); 2) an extended Resource Description Framework (RDF) triple attribute model that represents entity properties as independent graph nodes connected by semantic edges, enabling property-level querying and meta-reasoning; 3) coordinate-free spatial reasoning that leverages the Korean road numbering standard (20m interval, k = 10 m per unit) for proximity ranking without Global Positioning System (GPS) coordinates; and 4) a 14-strategy reasoning router that classifies queries into specialized graph traversal algorithms. Evaluated on a real-world dataset of Daejeon, South Korea (58,882 nodes, 80,088 edges from 58,408 buildings and 711 roads), GNLM achieves 88.4% accuracy on 250 diverse queries with a hallucination rate (HR) of only 5.7% on 35 fact-verification queries, compared to 60.0% for pure LLM, 25.7% for Dense RAG, and 17.1% for GraphRAG - demonstrating a consistent improvement as structural grounding increases. Cross-model validation with llama3.1:8b confirms LLM-agnostic generalizability. Multi-region validation on Sejong City (17,008 buildings) with identical methodology confirms cross-region generalizability (Yuseong-gu avg ρ = 0.976 , Sejong avg ρ = 0.973 , both over 10 roads). Road-number-based proximity ranking achieves Spearman ρ = 0.9878 and Precision@10= 96.0% against GPS ground truth, with short-range (<500m) Mean Absolute Error (MAE) of only 31.9m, demonstrating that structured address numbers alone can effectively preserve spatial ordering without any coordinate data. Cross-road queries, supported via a graph-native intersection topology with 34 annotated intersections, achieve a 90% resolution rate with Spearman ρ = 0.701 - substantially lower than same-road accuracy, indicating a current practical limitation that motivates the multi-hop and automated intersection-detection extensions.