TAAD: Teacher-Approved Arabic Diacritics

A small benchmark of 110 Arabic minimal pairs in which the diacritics alone decide the meaning. Each item is one sentence written two ways. Strip the short vowel marks and the two spellings become the identical string, so a model that cannot read the marks has no information with which to choose.

What is in it

Each item gives two vocalizations of the same consonantal skeleton, and one short continuation that fits each reading:

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

Both sentences reduce to أحب الشعر once diacritics are removed.

The intended use is a forced choice: given one reading of the sentence, is its own continuation scored above the other one? Each item is asked in both directions, giving 220 decisions in total, so a model that simply prefers one continuation wins once and loses once and gains nothing.

Construction constraints

Every item satisfies all of the following, checked mechanically:

1. The two sentences are identical once diacritics are stripped.
2. Shadda (U+0651) is identical in both readings. The stripping convention preserves shadda, so a pair differing in it would be visible to a diacritic-blind model.
3. Exactly one word differs between the readings, and it is the last word.
4. Both continuations carry the same final short vowel. This is the most important constraint. Without it, a model can succeed by copying the case vowel of the preceding word rather than by reading the ambiguous word. In an earlier version of this design that lacked the constraint, the copying shortcut alone resolved every position in the set.
5. The two continuations are different words, never one adjective in two cases.
6. One word per continuation.
7. The contrast is lexical, not a case contrast: the two forms differ inside the word, not only in a final inflectional vowel.

Provenance

Items were written and reviewed by an Arabic language teacher. Rows marked authored were composed by the reviewer; rows marked validated were drafted as candidates against the constraints above and then checked and approved by the same reviewer. All 110 passed review.

Candidates that failed any construction constraint were discarded before review.

Dataset structure

field	type	description
id	string	Stable item identifier, taad-001 … taad-110
sentence_a	string	Reading A, fully diacritized
ending_a	string	The one-word continuation that fits reading A
sentence_b	string	Reading B, fully diacritized
ending_b	string	The one-word continuation that fits reading B
word_pair	string	The ambiguous word, undiacritized
provenance	string	authored or validated
undiacritized	string	The shared consonantal string both readings reduce to

110 rows, one configuration, no splits.

Usage

from datasets import load_dataset

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

# the two forced choices for this item
#   given sentence_a: is ending_a scored above ending_b?
#   given sentence_b: is ending_b scored above ending_a?
print(item["sentence_a"], "->", item["ending_a"])
print(item["sentence_b"], "->", item["ending_b"])
print(item["undiacritized"])   # identical for both readings

Suggested scoring

Tokenize sentence + " " + continuation as a single string, run one forward pass, and sum the log-probability of the continuation's tokens only. The sentence is identical across the two continuations and would otherwise dominate the comparison.

Because the two continuations are different words, a raw sum favours the shorter one. Dividing by the character length of the undiacritized continuation gives a divisor that is a property of the item and identical for every model, so it cannot favour one tokenizer over another. Dividing by a model's own token count does not have this property and is not comparable across tokenizers.

A graded alternative that needs no normalization at all:

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, so its length, frequency and token count cancel exactly. What remains reflects only the change between the two vocalizations. Zero means the diacritics changed nothing.

A note on the expected floor

A model whose vocabulary cannot encode short vowels maps both readings of every item to the identical token sequence. Its two scores are then equal by construction and it scores exactly chance with exactly zero displacement. This is arithmetic rather than a measurement of ability, and it is the point of the set: these sentences are unresolvable without the marks as a matter of what is written.

When reporting confidence intervals, resample items rather than decisions. The two directions of an item are strongly anti-correlated by construction and are not independent observations.

Citation

@misc{taad2026,
  title  = {{TAAD}: Teacher-Approved Arabic Diacritics},
  author = {{Author names withheld during review}},
  year   = {2026},
  publisher = {Hugging Face},
  note   = {Author list and URL will be added once the accompanying paper is published}
}

License

CC BY 4.0.

Authors

Author names withheld during review.

Contact

Contact details will be added once the accompanying paper is published.