Uncertainty-driven Partial Diacritization for Arabic Text

What it is about

Full diacritization makes Arabic text dense, so the paper asks which diacritics to keep and which to drop. It uses a character-based encoder-decoder diacritic restorer (CATT) and scores each character with one of three uncertainty measures: softmax response, BALD (Monte Carlo dropout) and Mahalanobis distance. A character is kept or dropped by thresholding its score. On the Tashkeela test set, softmax response tracks the diacritic error rate best and is cheap to compute; keeping 80% of the diacritics gives a diacritic error rate of about 2.5%. On the out-of-domain ArVoice test set the base error rate is above 10%, and a Confident Error Regularizer trades recall for precision in finding error-prone words, at a cost in overall error. The authors call the work exploratory: there is no standard evaluation for partial diacritization yet.

Ideas

• I1. Use the softmax-response score of a diacritic model as a per-character keep/drop rule. A Hebrew analogue is the point set of unpointed-to-pointed text in tools/hbo: keep the vowel points the model is unsure of and drop the ones it predicts with confidence, and report a diacritic-error-versus-coverage curve.
• I2. Use the same confidence signal as a fallback switch: when the pointed model is unsure about a word, use the unpointed mode of tools/hbo for that word instead of guessing. The paper's numbers are the reference for how much this should cost in error.
• I3. Use the diacritic-error-versus-coverage curve as a measure: sweep the threshold and plot the error against the share of marks kept. This is a cheap, model-independent test to run on any pointing rule set, including the vowel rules in 1. Script and pointing.

For Mova

I1 and I2 apply to tools/hbo (the pointed and unpointed modes) and to the choice of which marks a reader needs, 1. Script and pointing and 2. Cantillation: the accents (te'amim). I3 is a measurement for the same tool. The paper's Confident Error Regularizer is a neural training method, so it does not transfer directly to Mova's rule and table components.

Cite

The paper itself, not this card. BibTeX from the library's record of the work. DOI 10.18653/v1/2025.uncertainlp-main.1

@inproceedings{alblooshi2025uncertainty,
  title = {{Uncertainty-driven Partial Diacritization for Arabic Text}},
  author = {Alblooshi, Humaid Ali and Shelmanov, Artem and Aldarmaki, Hanan},
  booktitle = {Proceedings of the 2nd Workshop on Uncertainty-Aware NLP (UncertaiNLP 2025)},
  year = {2025},
  doi = {10.18653/v1/2025.uncertainlp-main.1},
  url = {https://doi.org/10.18653/v1/2025.uncertainlp-main.1}
}

References

1 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Sawsan Alqahtani, Hanan Aldarmaki, and Mona Diab. 2019. Homograph disambiguation through selective diacritic restoration. In Proceedings of the Fourth Arabic Natural Language Processing Workshop, pages 49–59, Florence, Italy. Association for Computational Linguistics. find

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/aclbulk-2025-alblooshi-uncertainty-driven-partial-diacritization-arabic/