OtoBERT: Identifying Suffixed Verbal Forms in Modern Hebrew Literature

What it is about

Modern Hebrew literature sometimes fuses a verb with its direct-object pronoun into one word. The fused form is often identical to a common verb with another meaning, so readers are misled until the sentence ends. Annotators found 222 suffixed verbs in 135K words of literature (17.76 per 10K) against 7 in 185K words of news and 2 in 43K words of another news source; 64 of the literary cases were ambiguous. The authors pretrain a BERT-base model from scratch with a 128K vocabulary in which each verb-plus-pronoun pair is one token. A form is flagged when at least N of the top K masked-language-model predictions are such tokens. Tested on 2,589 unsuffixed and 264 suffixed homographs, K = 1000 gives precision 100% at N = 25 (recall 19%), 84% at N = 10 (recall 52%) and 54% at N = 5 (recall 74%). Fine-tuned classifiers on mBERT, AlephBERT and DictaBERT reach F1 0.39–0.58, and the DictaBERT morphological tagger has recall of 24%. The model and dataset are public; the model is Apache-licensed as stated in the paper.

Ideas

• I1. Paired readings in the lexicon: for each ambiguous form, store both analyses (a bare verb, and a verb with a person-gender object clitic) with the clitic's features. The paper's examples ("they ordered" against "he ordered it") make good seed pairs. In UD this is a verb with a pronoun object merged in the surface form, so the segmenter has to split it.
• I2. Confidence tiers with a stated precision: the paper's thresholds give three tiers (precision 100%, 84%, 54%), each with a fixed hedge in the gloss ("likely", "perhaps"). Mova can output the same three tiers with the rule that fired, so the gloss says why.
• I3. A held-out test: the 264 suffixed and 2,589 unsuffixed homographs are a clean test for the Modern Hebrew parser, and the paper's own tagger result (recall 24%) is a baseline to beat. Check the dataset's licence before use, because the text does not state it.

For Mova

I1 goes into 05b. Verb paradigms: the strong verb in seven binyanim and the weak roots (paired readings) and 05. The verb: system, binyanim, tense and use (object clitics). I2 is an explanation rule in prag. I3 is a test for 11. Modern Hebrew in Universal Dependencies.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2024.tsar-1.2

@inproceedings{shmidman2024otobert,
  title = {{OtoBERT: Identifying Suffixed Verbal Forms in Modern Hebrew Literature}},
  author = {Shmidman, Avi and Shmidman, Shaltiel},
  booktitle = {Proceedings of the Third Workshop on Text Simplification, Accessibility and Readability (TSAR 2024)},
  year = {2024},
  doi = {10.18653/v1/2024.tsar-1.2},
  url = {https://doi.org/10.18653/v1/2024.tsar-1.2}
}

References

10 references, 3 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. S. Y. Agnon. 1953a. At the Handles of the Lock. Schocken Publishing House. find
2. S. Y. Agnon. 1953b. Only Yesterday. Schocken Publishing House. find
3. Tuqa Bani Yaseen, Qusai Ismail, Sarah Al-Omari, Eslam Al-Sobh, and Malak Abdullah. 2021. JUST-BLUE at SemEval-2021 task 1: Predicting lexical complexity using BERT and RoBERTa pre-trained language models. In Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 661–666, Online. Association for Computational Linguistics. find
4. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019a. Bert: Pre-training of deep bidirectional transformers for language understanding. Preprint, arXiv:1810.04805. ↗
5. Shaltiel Shmidman, Avi Shmidman, Moshe Koppel, and Reut Tsarfaty. 2024. MRL parsing without tears: The case of Hebrew. In Findings of the Association for Computational Linguistics ACL 2024, pages 4537–4550, Bangkok, Thailand and virtual meeting. Association for Computational Linguistics. find our card
6. Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019b. Bert: Pre-training of deep bidirectional transformers for language understanding. Preprint, arXiv:1810.04805. ↗
7. Xinying Song, Alex Salcianu, Yang Song, Dave Dopson, and Denny Zhou. 2021. Fast wordpiece tokenization. Preprint, arXiv:2012.15524. ↗
8. Eylon Gueta, Avi Shmidman, Shaltiel Shmidman, Cheyn Shmuel Shmidman, Joshua Guedalia, Moshe Koppel, Dan Bareket, Amit Seker, and Reut Tsarfaty. 2023. Large pre-trained models with extra-large vocabularies: A contrastive analysis of hebrew bert models and a new one to outperform them all. Preprint, arXiv:2211.15199. ↗ our card
9. Reut Tsarfaty and Yoav Goldberg. 2008. Word-based or morpheme-based? annotation strategies for Modern Hebrew clitics. In Proceedings of the Sixth International Conference on Language Resources and Evaluation (LREC’08), Marrakech, Morocco. European Language Resources Association (ELRA). find our card
10. Abdelhak Kelious, Mathieu Constant, and Christophe Coeur. 2024. Complex word identification: A comparative study between ChatGPT and a dedicated model for this task. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), pages 3645–3653, Torino, Italia. ELRA and ICCL. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2024-shmidman-otobert-identifying-suffixed-verbal-forms/