The Potential of Grounded Synthetic Corpora to Address Persistent Geoparsing Challenges

Geoparsing, extracting and geocoding place mentions from text, remains geographically and linguistically uneven, with persistent under-representation of non-English contexts and less prominent places. Manual corpus annotation is prohibitively expensive, and existing automated approaches are limited in their annotation completeness. This paper argues that grounded synthetic corpora, generated by LLMs but anchored in gazetteers and contextual sources such as Wikipedia, offer a scalable complement to human-authored benchmarks. By inverting the conventional workflow through sampling places first, then generating naturalistic text around them toponyms become pre-disambiguated by construction, enabling balanced spatial coverage and multilingual scope. We outline the benefits, identify new bias pathways, and set out the key research challenges for producing gold-standard synthetic corpora that reduce bias without introducing new ones.