Balancing factoid retrieval and candidate recommendation to identify toponym origins in gazetteers

This paper investigates the adaptation of large language models (LLMs) for gazetteer enrichment, specifically to resolve missing placename origins in the Melbourne gazetteer (Australia). Gazetteers are foundational databases for geographic information systems. However, querying toponym etymologies remains challenging: populating such attributes resists rule-based inference and necessitates direct semantic retrieval from unstructured resources. We adopted two search strategies: factoid retrieval for explicit origins, and candidate recommendation otherwise. The use of nonspatial models risks degraded spatial matching performance. Hence, we developed a retrieval-augmented generation (RAG) architecture within a geographic question-answering framework. Our system employed parallel semantic-spatial rankers that deliberately enforced a geographic bias toward target regions. Our system improved spatial ranking scores by up to 7% in normalized discounted cumulative gain. However, at the final stage, LLMs still struggled to reliably filter spatial matches. To investigate this limitation, we conducted a probing experiment on standalone LLMs. We revealed intrinsic spatial distortions regarding adjacency relations in Greater Melbourne, notably an inherent popularity bias that prevented fine-grained responses. Our findings suggest that effectively deploying LLMs for spatial data origins must advance beyond factoid retrieval by also refining candidate recommendation at every stage: calibrated biases adapt language model search to local specificities.