khalidmokha/lexibert-ud-uralic

LexiBERT UD parser for Uralic languages

One model for UPOS, UFeats, lemmas and dependency trees, trained on the training splits of all Uralic UD treebanks together with 6,000 Finnish TDT and 4,000 Estonian EDT sentences. It is LexiBERT with a learned mix of the encoder layers, classifiers for UPOS, UFeats and lemma edit rules, and a biaffine parser; the tree is the highest scoring one with one root. The input must be tokenized. The relations are given without subtypes.

pip install "lexibert @ git+https://github.com/mokha/lexibert"
lexibert-parse --model khalidmokha/lexibert-ud-uralic --fst myv < input.conllu > output.conllu

from lexibert import Parser
from lexibert.parser import to_conllu

parser = Parser.from_pretrained("khalidmokha/lexibert-ud-uralic")
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, if the FST knows it (pip install uralicNLP and uralicApi.download("myv")).

Results

UD test sets, gold tokenization. Lem_FST: the lemma restricted to the FST lemmas.

Treebank	UPOS	UFeats	Lem_FST	UAS	LAS
Erzya JR	90.9	75.0	93.8	80.1	71.0
Moksha JR	81.7	58.8	92.5	73.3	61.3
Komi-Zyrian Lattice	86.2	54.0	92.5	77.4	66.7
Komi-Zyrian IKDP	80.6	52.8	85.9	67.6	53.0
Komi-Permyak UH	85.7	51.9	91.7	83.7	73.9
Skolt Sámi Giellagas	67.8	53.9	92.8	52.7	37.6
Livvi KKPP	86.8	63.6	85.4	78.0	66.0
Karelian KKPP	91.3	68.5	91.6	74.5	65.9
North Sámi Giella	94.2	92.2	96.1	80.7	76.0
Veps VWT	87.6	71.6	71.3	81.3	69.5
Nenets Tundra	36.4	63.1	34.0	38.4	13.5
Average	80.9	64.2	84.3	71.6	59.5

Most of these treebanks have no training data of their own. For Nenets, the lemmas of the treebank do not follow the conventions of the FST.

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}
}

The parsers trained on Erzya or North Sámi only (3 seeds each): lexibert-ud-single-treebank.

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