DictaBERT-splinter: Splintering Nonconcatenative Languages for Better Tokenization

DictaBERT-splinter is a BERT-style language model for Hebrew, released here.

This is the base model pretrained with the masked-language-modeling objective.

Sample usage:

from transformers import AutoModelForMaskedLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained('dicta-il/dictabert-splinter', trust_remote_code=True)
model = AutoModelForMaskedLM.from_pretrained('dicta-il/dictabert-splinter')

model.eval()

sentence = 'בשנת 1948 השלים אפרים קישון את [MASK] בפיסול מתכת ובתולדות האמנות והחל לפרסם מאמרים הומוריסטיים'

output = model(tokenizer.encode(sentence, return_tensors='pt'))
# the [MASK] is the 7th token (including [CLS])
import torch
top_2 = torch.topk(output.logits[0, 7, :], 2)[1]
print('\n'.join(tokenizer.batch_decode(top_2))) # should print התמחותו / לימודיו 

Citation

If you use DictaBERT-splinter in your research, please cite Splintering Nonconcatenative Languages for Better Tokenization

BibTeX:

@misc{gazit2025splinteringnonconcatenativelanguagesbetter,
      title={Splintering Nonconcatenative Languages for Better Tokenization}, 
      author={Bar Gazit and Shaltiel Shmidman and Avi Shmidman and Yuval Pinter},
      year={2025},
      eprint={2503.14433},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.14433}, 
}

License

Shield:

This work is licensed under a Creative Commons Attribution 4.0 International License.