We argue that languages are not alike in terms of their structure and the availability of (labelled) data. For example, Finnish place names may need to be lemmatised (Helsingissä -> Helsinki) to query databases effectively.

Check out this geoparsing pipeline of three tasks and three possible outcomes for each place name. Note that lemmatisation errors are a major source of noise and failures.

To tackle geoparsing for Finnish, we draw from existing resources where possible. We fine-tune a Finnish language model with a modified named entity corpus and set up an instance of a Pelias geocoder.

As for the resources that are missing, such as labelled Finnish corpus for evaluating geoparsers? We annotated tweets and news articles – both corpora are shared openly.

Finally, the big question: does it work? Yeah, kinda. Despite the simple architecture, the geoparser performed well enough on the two corpora.

If you want to check out the geoparser: https://github.com/DigitalGeographyLab/Finger-geoparser
I also put out a blog post: https://blogs.helsinki.fi/digital-geography/2024/07/05/new-article-out-geographical-and-linguistic-perspectives-on-developing-geoparsers-with-generic-resources/