The construction of annotated geoparsing corpora is costly and has resulted in limited linguistic and geographic coverage, particularly outside English-speaking regions. This paper investigates the use of Large Language Models (LLMs) to generate synthetic geoparsing corpora for four languages in four regions: Austria (German), Belarus (Belarusian), Finland (Finnish), and Ghana (English). We evaluate the generated corpora through automatic and human quality checks, and by benchmarking state-of-the-art geoparsers on the synthetic data. Our results show that while LLM-generated corpora enable credible geoparser evaluation, low-resource regions and languages expose systematic limitations in LLM-powered synthetic text generation approaches related to underlying geographic data coverage, completeness, and linguistic variation. We make our code available on GitHub.

Peer reviewed