Assessing Spatial Accuracy of Nature Observations Using the LLM-Based Geoparser Service

Geo-parsing tools have historically relied on AI-based techniques for tasks such as text processing and toponym disambiguation. However, the rapid proliferation of large language model based tools and services necessitates a re-evaluation of traditional geoparser architectures to explore performance enhancements, introduce new functionalities, support different languages, or allow for rapid prototyping and development of new geoparser-based services. This paper focuses on the functional components of typical geoparsers and examines the effects of substituting one or more of these elements with LLM-based functionality. The performance impact of these modifications is evaluated in the context of a developed online service for assessing the spatial accuracy of nature observations. The performance of the service is compared against the baseline performance of the open-source geoparser, Mordecai3, using datasets in Croatian language and with varying complexity of location descriptions.