SpatialKG
收藏资源简介:
SpatialKG是一个高质量的地理空间知识图谱,基于OpenStreetMap和Foursquare的开放源城市数据源构建。该图谱包含了城市环境中基本的实体类型和它们之间的关系,例如兴趣点(POI)、兴趣区域(AOI)和道路。SpatialKG用于构建GEOHALUBENCH,这是一个包含3个一级类别和5个二级类别的地理空间幻觉基准,旨在评估和缓解LLM中的地理空间知识幻觉。GEOHALUBENCH包含了2100个测试样例,每个样例都包含一个多选题,一个参考答案,以及一个将选项映射到幻觉类型的映射。
SpatialKG is a high-quality geospatial knowledge graph constructed based on open-source urban data sources from OpenStreetMap and Foursquare. This knowledge graph includes basic entity types and their interrelationships in urban environments, such as Points of Interest (POIs), Areas of Interest (AOIs) and roads. SpatialKG is used to develop GEOHALUBENCH, a geospatial hallucination benchmark that consists of 3 primary categories and 5 secondary categories, aiming to evaluate and alleviate geospatial knowledge hallucinations in Large Language Models (LLMs). GEOHALUBENCH contains 2100 test samples, each of which comprises a multiple-choice question, a reference answer, and a mapping that links options to hallucination types.

- 1Mitigating Geospatial Knowledge Hallucination in Large Language Models: Benchmarking and Dynamic Factuality Aligning清华大学 · 2025年



