eFontes Latin Tagger v2 — lemmatization model
This model lemmatises Latin token forms. It is the lemma component of the corrected eFontes Latin Tagger v2 release.
Intended use
The model interface is FORM:POS to LEMMA. For normal use, install the
public eFontes Tagger runtime,
which supplies POS labels, preserves punctuation and numeral forms directly,
and returns annotation output without a separate cleanup script.
Evaluation
On the eFontes held-out test split (18,821 tokens), this snapshot achieved 98.48% exact lemma accuracy. Raw output contained 0 POS-label artefacts. The public runtime handles punctuation and unambiguous numeral forms deterministically, rather than sending those non-lexical forms to the generator.
Limitations
The model is intended for Medieval Latin. Accuracy can differ for periods, orthographies, abbreviations, damaged text, and genres not represented in its training and evaluation data. Predictions require scholarly review.
Authorship and lineage
The corrected v2 release is maintained and published by Krzysztof Nowak (2026). It builds on the earlier eFontes research and models described in the 2024 paper below.
Citation
Please cite the versioned eFontes Tagger v2 release together with the original eFontes paper:
Nowak, Krzysztof; Jędrzej Ziębura; Krzysztof Wróbel; and Aleksander Smywiński-Pohl. 2024. eFontes. Part of Speech Tagging and Lemmatization of Medieval Latin Texts: A Cross-Genre Survey. https://doi.org/10.48550/arXiv.2407.00418
Licence
Model weights are released under CC BY-NC-SA 4.0.