geo-distiluse-swissnames3d

Sentence-transformer for toponym resolution against the swissNAMES3D gazetteer, for use with the Irchel Geoparser. It follows the approach of dguzh/geo-all-MiniLM-L6-v2, fine-tuned from distiluse-base-multilingual-cased-v1 on German-language Swiss news articles.

Trained on the default candidate descriptions of the Irchel Geoparser (name, class, municipality, district, canton).

Usage

from geoparser.modules import SentenceTransformerResolver

resolver = SentenceTransformerResolver(
    model_name="segmue/geo-distiluse-swissnames3d",
    gazetteer_name="swissnames3d",
)

Variants

Model	Thesis	Candidate descriptions
geo-distiluse-swissnames3d	M3	default
geo-distiluse-swissnames3d-config1	M4	spatial context, config1
geo-distiluse-swissnames3d-config2	M5	spatial context, config2

Training

8,255 text–description pairs, contrastive loss, learning rate 1e-5, 2 epochs. Training and evaluation code: ma-experiments.

Source

Master's thesis, University of Zurich, 2026: Describing Places by Their Surroundings: Enriching Candidate Descriptions with Spatial Context for Toponym Resolution.