GANDR - Georelating Dataset, Metrics, and Evaluation
Abstract: Georelating has been introduced to learn geospatial representations of events from textual reports, which requires the interpretation of spatial relations. To foster the development and evaluation of Georelating systems, we construct the silver-standard Georelating Annotated Natural Disaster Reports dataset GANDR and benchmark our LLM agent architecture as a baseline (areal F1 = 0.609, fuzzy cell match score = 0.833) for this new task. GANDR comprises synthetic disaster reports referencing 1,000 US and 1,000 EU cities, annotated with Discrete Global Grid System (DGGS) cells for efficient geospatial integration. We propose a set of five complementary metrics capitalizing on the DGGS annotations for efficient and comprehensive evaluation. Analysis reveals the potential of reasoning LLMs integrated with geographical knowledge bases to address variation across spatial relations such as (inter-)cardinal directions. We highlight the estimation of the impact area’s size as a key challenge of Georelating.
Show BibTeX
@inproceedings{DBLP:conf/geoai/MoltzenU25,
author = {Kai Moltzen and
Ricardo Usbeck},
editor = {Shawn D. Newsam and
Lexie Yang and
Song Gao and
Di Zhu},
title = {{GANDR} - Georelating Dataset, Metrics, and Evaluation},
booktitle = {Proceedings of the 8th {ACM} {SIGSPATIAL} International Workshop on
{AI} for Geographic Knowledge Discovery, GeoAI 2025, The Graduate
Hotel Minneapolis, Minneapolis, MN, USA, November 3-6, 2025},
pages = {61--71},
publisher = {{ACM}},
year = {2025},
url = {https://doi.org/10.1145/3764912.3770819},
doi = {10.1145/3764912.3770819},
timestamp = {Wed, 25 Feb 2026 08:25:21 +0100},
biburl = {https://dblp.org/rec/conf/geoai/MoltzenU25.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}
}