khalidmokha/lexibert-sentiment-sst2

LexiBERT sentiment (trained on English only)

LexiBERT fine-tuned on English SST-2 for 2,000 steps (seed 1), with a linear layer on the average of the outputs. It has never seen Uralic training data for this task.

from lexibert import SentimentClassifier

clf = SentimentClassifier.from_pretrained("khalidmokha/lexibert-sentiment-sst2")
clf.predict(["..."])   # ['positive'] or ['negative']

Results

Accuracy of this model on the SST-2 development data and on the aligned test data of Alnajjar et al. (2023):

SST-2 dev	Finnish	Komi-Zyrian	Moksha	Erzya	Udmurt
91.6	76.5	76.5	66.2	67.6	64.7

The paper reports the mean of three seeds (the other two are in the Zenodo record). For Erzya, Moksha, Komi-Zyrian and Udmurt, the results improve clearly when the sentences are first glossed into English word by word with the dictionary (scripts/26_sentiment_gloss.py in the code).

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 models and seeds: https://doi.org/10.5281/zenodo.23545486.