Automatic Text Segmentation of Ancient and Historic Hebrew

What it is about

Unpunctuated Rabbinic and historic Hebrew and Aramaic texts run to hundreds of words without a full stop, and most neural tools need input cut into chunks. The authors annotate a new ground-truth set of 25 texts (7,265 words, one annotator) with three marker levels: Break (B), Partial (P), and Maybe (M), where M positions are clearly better than random but should not be cut. They compare segmenters: few-shot completion with the DictaLM 2.0 base model, masked-language-model prediction of punctuation with several Hebrew BERT models (HeBERT, AlephBERT, DictaBERT, BEREL), the multilingual Segment-Any-Text model, and instruction-prompted GPT-4o and Claude Sonnet. The aim is chunks of at most about 50 words. BEREL, trained on Rabbinic Hebrew, is the best masked-model option and rivals the LLMs without their instability; Segment-Any-Text does well but trails BEREL.

Ideas

• I1. A rule-based chunker before parsing: cut first at the B positions, then at P positions when a chunk would pass 50 words, and never at M positions unless nothing else is possible. Each cut then has a reason, and the rule can be run on tools/hbo input before gparse.
• I2. Score segmenters at three strictness levels (B; B+P; B+P+M), as the paper does, so that a rule is judged both on precision and on how often it breaks a sentence badly.
• I3. Use the 25-text set as a held-out test for the chunker, with the 50-word cap as a hard check. Run it once per release, not during tuning.
• I4. Cut positions as clause-boundary evidence: the B and P markers are the same kind of boundary that the parser's clause rules need, so a chunker that misses them is a signal for Syntax: the clause and the sentence.

For Mova

I1 to I3 belong in tools/hbo, in the segmentation step before gparse, with the test set kept held out. I4 is a note for Syntax: the clause and the sentence. The paper's best system is a masked model; Mova should take the measurement protocol and the task, and build the rule, not the model.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.alp-1.1

@inproceedings{rosensweig2025automatic,
  title = {{Automatic Text Segmentation of Ancient and Historic Hebrew}},
  author = {Rosensweig, Elisha and Resnick, Benjamin and Gershuni, Hillel and Guedalia, Joshua and Dershowitz, Nachum and Shmidman, Avi},
  booktitle = {Proceedings of the Second Workshop on Ancient Language Processing},
  year = {2025},
  doi = {10.18653/v1/2025.alp-1.1},
  url = {https://doi.org/10.18653/v1/2025.alp-1.1}
}

References

12 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Avihay Chriqui and Inbal Yahav. 2021. HeBERT & HebEMO: a Hebrew BERT Model and a Tool for Polarity Analysis and Emotion Recognition. Preprint, arXiv:2102.01909. ↗ our card
2. Aitor Alvarez, Carlos-D Martínez-Hinarejos, Haritz Arzelus, Marina Balenciaga, and Arantza del Pozo. 2017. Improving the automatic segmentation of subtitles through conditional random field. Speech Communication, 88:83–95. find
3. Eep Talstra Centre for Bible and Computer. 2023. Bhsa: Biblia hebraica stuttgartensia amstelodamen-sis. https://github.com/ETCBC/bhsa. Accessed: 2025-03-18. ↗
4. Dennis Aumiller, Satya Almasian, Sebastian Lackner, and Michael Gertz. 2021. Structural text segmentation of legal documents. In Proceedings of the Eighteenth International Conference on Artificial Intelligence and Law, pages 2–11. find
5. Federico Errica, Giuseppe Siracusano, Davide Sanvito, and Roberto Bifulco. 2024. What Did I Do Wrong? Quantifying LLMs’ Sensitivity and Consistency to Prompt Engineering. Preprint, arXiv:2406.12334. ↗
6. Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam Mc-Candlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners. Preprint, arXiv:2005.14165. ↗
7. Markus Frohmann, Igor Sterner, Ivan Vulić, Benjamin Minixhofer, and Markus Schedl. 2024. "Segment Any Text: A Universal Approach for Robust, Efficient and Adaptable Sentence Segmentation". In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 11908– 11941, Miami, Florida, USA. Association for Computational Linguistics. find
8. Hongyu Gong, Yelong Shen, Dian Yu, Jianshu Chen, and Dong Yu. 2020. Recurrent chunking mechanisms for long-text machine reading comprehension. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 6751– 6761, Online. Association for Computational Linguistics. find
9. Jesús Calleja, Thierry Etchegoyhen, and David Ponce. 2024. Automating Easy Read Text Segmentation. Preprint, arXiv:2406.11464. ↗
10. Sangwoo Cho, Kaiqiang Song, Xiaoyang Wang, Fei Liu, and Dong Yu. 2022. find
11. Toward Unifying Text Amir Hazem, Béatrice Daille, Dominique Stutzmann, Christopher Kermorvant, and Louis Chevalier. 2020. Hierarchical text segmentation for medieval manuscripts. In Proceedings of the 28th International Conference on Computational Linguistics, pages 6240–6251. find
12. A Timo Homburg and Christian Chiarcos. 2016. Akkadian word segmentation. In Proceedings Tenth International Conference on Language Resource Evaluation.(LREC 2016), pages 4067–4074. A.1 find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2025-rosensweig-automatic-text-segmentation-ancient-historic/