khalidmokha/lexibert-ud-single-treebank

LexiBERT UD parsers for Erzya and North Sámi

UD parsers that were trained on a single treebank, Erzya JR or North Sámi Giella, with three random seeds each. They predict UPOS, UFeats, the lemma and the dependency tree (relations without subtypes) on tokenized text. The encoder is LexiBERT. For all of the Uralic treebanks, use the multilingual parser lexibert-ud-uralic instead.

Folder	Trained on	Seed
myv-seed1, myv-seed2, myv-seed3	Erzya JR	1, 2, 3
sme-seed1, sme-seed2, sme-seed3	North Sámi Giella	1, 2, 3

Use

With the lexibert package (code):

pip install "lexibert @ git+https://github.com/mokha/lexibert" uralicNLP

from huggingface_hub import snapshot_download
from lexibert import Parser
from lexibert.parser import to_conllu

folder = snapshot_download("khalidmokha/lexibert-ud-single-treebank", allow_patterns=["myv-seed1/*"])
parser = Parser.from_pretrained(f"{folder}/myv-seed1")
print(to_conllu(parser.parse([["Мон", "вечкан", "кудом", "."]], fst_lang="myv")))

With fst_lang, the lemma must be one of the lemmas that the UralicNLP FST gives for the word, when the FST knows it.

Results

On the test sets, on gold tokenization, mean of the three seeds:

Treebank	UPOS	UFeats	Lemma	Lemma (FST)	UAS	LAS
Erzya JR	90.8	73.9	81.9	93.8	74.9	66.0
North Sámi Giella	94.2	91.2	80.9	95.9	77.7	73.0

Citation

@inproceedings{alnajjar-etal-2026-grounding,
  title     = {Grounding Encoders and Word Embeddings in {FST}s and Dictionaries for {U}ralic Languages},
  author    = {Alnajjar, Khalid and H{\"a}m{\"a}l{\"a}inen, Mika and Rueter, Jack and Partanen, Niko},
  booktitle = {Proceedings of the International Workshop on Computational Linguistics for Uralic Languages (IWCLUL 2026)},
  year      = {2026}
}

Code: https://github.com/mokha/lexibert · All UD parsers are also in the Zenodo record https://doi.org/10.5281/zenodo.23545487.