Knowledge-Enhanced Geospatial QA: Integrating Wikidata Fact Verification with LLMs

Abstract

Geospatial question answering (QA) challenges large language models (LLMs) in reasoning about geospatial relationships. This paper presents a hybrid framework that integrates LLMs with Wikidata for fact verification and retrieval, enhancing their geospatial reasoning capabilities. The framework generates facts, verifies them against Wikidata, and uses validated knowledge for use in a Retrieval-Augmented Generation (RAG) pipeline. Experimental results demonstrate that this approach outperforms zero-shot prompting for all tested models, including GPT-3.5-turbo-0125, Llama-3-8b, and Qwen-2.5-14b, showcasing its effectiveness in improving geospatial QA accuracy.