TAAD — Teacher-Approved Arabic Diacritics

TAAD is a small evaluation benchmark of 110 Arabic minimal pairs in which the diacritics alone determine the intended meaning.

Each item contains the same underlying consonantal string with two different vocalizations. Once the relevant Arabic short-vowel marks are removed, the two readings become identical. A model therefore needs access to the diacritics to distinguish the intended interpretations.

TAAD was developed during the evaluation and public review process surrounding NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic, published in Transactions on Machine Learning Research (TMLR) in 2026.

Related publication and public review discussion
Ali Issa, Tina Yaacoub, and Patrick Abi Salloum.
NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic.
Transactions on Machine Learning Research (TMLR), 2026.
https://openreview.net/forum?id=Vyc4Ormkqp

TAAD was discussed in the public OpenReview review and author-response thread associated with the NOHA paper, but it is not included as a benchmark in the final archival TMLR manuscript.

What is in it

Each item provides two vocalizations of the same consonantal skeleton together with one short continuation appropriate for each reading.

	Arabic	Reading
A	أُحِبُّ الشِّعْرَ → العَرَبِيَّ	I love Arabic poetry
B	أُحِبُّ الشَّعْرَ → الطَّوِيلَ	I love long hair

Both sentence prefixes reduce to:

أحب الشعر

once the relevant short-vowel diacritics are removed.

The intended use is a paired forced-choice evaluation:

• given sentence_a, determine whether ending_a is scored above ending_b;
• given sentence_b, determine whether ending_b is scored above ending_a.

Each of the 110 items is therefore evaluated in both directions, yielding 220 forced-choice decisions.

Because each continuation appears once as the intended continuation and once as the competing continuation, a fixed preference for one continuation cannot by itself improve performance across the pair.

Construction constraints

Every released item satisfies the following constraints, checked mechanically:

1. The two sentence readings are identical after the relevant short-vowel diacritics are stripped.

2. Shadda (U+0651) is identical in both readings.
   The stripping convention preserves shadda, so a pair differing in shadda would remain distinguishable even after the other short-vowel marks were removed.

3. Exactly one word differs between the two vocalizations, and it is the final word of the sentence prefix.

4. Both continuations carry the same final short vowel.
   This prevents a model from solving the task merely by copying the case vowel of the preceding word instead of interpreting the ambiguous vocalized word.

5. The two continuations are different lexical items rather than the same word appearing in two different grammatical cases.

6. Each continuation consists of exactly one word.

7. The ambiguity is lexical rather than purely case-based: the two readings differ internally within the word and not only in a final inflectional vowel.

Provenance

The benchmark was designed and assembled by Ali Issa.

Items were written and reviewed with the involvement of an Arabic-language teacher.

Rows marked authored were composed directly by the teacher. Rows marked validated were first drafted as candidate items satisfying the benchmark's construction constraints and were subsequently reviewed and approved by the same teacher.

All 110 released items passed this review process.

Candidate items that violated any construction constraint were discarded before inclusion in the released dataset.

The benchmark was not scraped from an existing evaluation set and was not translated from another language.

Dataset structure

Field	Type	Description
id	string	Stable item identifier, from taad-001 to taad-110.
sentence_a	string	Reading A, fully diacritized.
ending_a	string	One-word continuation appropriate for reading A.
sentence_b	string	Reading B, fully diacritized.
ending_b	string	One-word continuation appropriate for reading B.
word_pair	string	Ambiguous word represented without the relevant short-vowel diacritics.
provenance	string	Either authored or validated.
undiacritized	string	Shared form obtained after stripping the relevant diacritics.

The dataset contains 110 rows and one configuration.

The Hugging Face dataset is exposed through a split named train for loading convenience. TAAD is an evaluation benchmark and is not intended as training data.

Usage

from datasets import load_dataset

ds = load_dataset("ali-issa/TAAD", split="train")
item = ds[0]

# Two paired forced-choice comparisons:
#
# sentence_a: ending_a should score above ending_b
# sentence_b: ending_b should score above ending_a

print(item["sentence_a"], "->", item["ending_a"])
print(item["sentence_b"], "->", item["ending_b"])
print(item["undiacritized"])

Suggested scoring

For each candidate continuation, tokenize

sentence + " " + continuation

as a single sequence and compute the log-probability of the continuation tokens only.

The sentence context should not itself be included in the scoring sum.

Length-normalized forced-choice score

Because the two candidate continuations may have different lengths, raw summed log-probability can favor shorter continuations.

For comparisons across tokenizers, one possible normalization is to divide the continuation log-probability by the character length of its undiacritized form.

Unlike normalization by a model's own token count, character length is a property of the text rather than of the tokenizer being evaluated.

Displacement score

A paired alternative that does not require length normalization is:

displacement =
    [logP(end_a | sent_a) - logP(end_b | sent_a)]
  - [logP(end_a | sent_b) - logP(end_b | sent_b)]

Each continuation appears once with each sign. This paired construction reduces fixed effects associated with the continuation itself, such as lexical frequency, length, or tokenization cost.

A displacement value of zero means that changing the vocalization did not change the model's relative preference between the two continuations.

Expected behavior without usable diacritics

If a model's input representation removes or cannot distinguish the relevant short-vowel diacritics, the two readings of an item collapse to the same input representation.

Such a model therefore has no information in the written input with which to systematically distinguish the two intended readings. Under the paired forced-choice construction, its expected performance is at chance, and its displacement is zero when the two contexts become representation-identical.

When estimating confidence intervals, resample items rather than individual directions. The two directions derived from the same minimal pair are paired observations and should not be treated as independent samples.

Relationship to the NOHA paper and review

TAAD was developed during the evaluation and review process for:

NOHA: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic
Ali Issa, Tina Yaacoub, and Patrick Abi Salloum
Transactions on Machine Learning Research (TMLR), 2026

https://openreview.net/forum?id=Vyc4Ormkqp

TAAD appears in the public OpenReview review and author-response discussion associated with the paper.

However, TAAD was not included as an evaluation benchmark in the final archival TMLR manuscript.

The NOHA paper and its OpenReview forum are therefore provided here as context for the research and review process in which TAAD was developed. The NOHA paper should not be interpreted as the publication formally introducing TAAD.

Citation

If you use TAAD in your work, please cite the dataset directly:

@misc{issa2026taad,
  title     = {{TAAD}: Teacher-Approved Arabic Diacritics},
  author    = {Ali Issa},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/ali-issa/TAAD},
  note      = {Arabic diacritic minimal-pair evaluation benchmark}
}

When discussing the broader research context in which TAAD was developed, the associated NOHA paper may additionally be cited:

@article{
issa2026noha,
title={{NOHA}: A Diacritic-Aware Rule-Based Morphophonological Tokenizer for Arabic},
author={Ali Issa and Tina Yaacoub and Patrick Abi Salloum},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2026},
url={https://openreview.net/forum?id=Vyc4Ormkqp},
note={}
}

License

TAAD is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

You may use, share, and adapt the dataset, including commercially, provided appropriate attribution is given.

https://creativecommons.org/licenses/by/4.0/

Author

Ali Issa
Université Saint-Joseph de Beyrouth

Acknowledgements

TAAD was developed in the broader context of the NOHA research project.

The author thanks the Arabic-language teacher who contributed to the authorship and validation of the benchmark items, and Tina Yaacoub for supervision and feedback on the broader research.

Contact

Questions, corrections, and reports concerning individual items are welcome.

Please open a discussion on this Hugging Face repository.