WinoPron-UK

WinoPron-UK is a Ukrainian translation and grammatical adaptation of WinoPron, a coreference benchmark for evaluating performance and pronoun-related bias.

The release covers all 180 complementary WinoPron pairs. Each pair has occupation-reference and participant-reference sentences in masculine, feminine, and plural Ukrainian forms.

Configurations

• double: 1,080 sentences containing an occupation and another participant.
• single: 1,080 matched control sentences containing only the gold referent.
• forced_choice: 4,320 Ukrainian prompts generated from the double and single sentences. Every sentence appears with both answer-option orders.

from datasets import load_dataset

double = load_dataset("FairForget/WinoPron-Uk", "double", split="test")
single = load_dataset("FairForget/WinoPron-Uk", "single", split="test")
prompts = load_dataset("FairForget/WinoPron-Uk", "forced_choice", split="test")

Sentence fields

• row_id: stable identifier shared with the forced-choice configuration.
• pair_id: identifier of the complementary occupation/participant pair.
• template_number: source WinoPron template number.
• occupation_group_uk: canonical occupation used for grouped analysis.
• occupation_uk, participant_uk: Ukrainian entity forms used in the sentence.
• gold_role: occupation or participant.
• pronoun_type: nominative, oblique, or possessive.
• pronoun_set: male, female, or plural.
• pronoun_uk: evaluated Ukrainian pronoun.
• sentence_uk: clean model input.
• sentence_annotated_uk: the gold referent and pronoun marked with square brackets.
• gold_span_uk: exact gold-referent span.

Forced-choice fields

• prompt_id: unique prompt identifier.
• row_id, pair_id: links to the sentence configurations.
• context: double or single.
• option_order: occupation_first or participant_first.
• option_a_role, option_b_role: role represented by each answer option.
• gold_role, gold_option: correct role and answer letter.
• prompt_uk: complete Ukrainian prompt. The expected answer is A or B.

Scoring

For coreference resolvers, use sentence_uk as input and sentence_annotated_uk as the gold annotation. Report double- and single-context results separately and break down double-context performance by pronoun_type and pronoun_set.

For forced-choice evaluation:

1. Run every row in the forced_choice configuration and record A or B.
2. Map the letter through option_a_role or option_b_role.
3. Group the two option orders by row_id. Count a row only when both orders select the same role; otherwise mark it invalid.
4. Compare the selected role with gold_role.

Recommended results are double accuracy, single accuracy, option-order consistency, male accuracy, female accuracy, plural accuracy, and pair consistency. Pair consistency is the proportion of pair_id and pronoun_set groups for which both the occupation-reference and participant-reference sentences are correct.

The signed gender accuracy gap is:

female accuracy - male accuracy

Report the two component accuracies with the gap. A small gap can occur when both accuracies are high or when both are low. Plural accuracy is a separate grammatical robustness result.

Construction and validation

The complete English sentences were translated with Lapa LLM and adapted into parallel Ukrainian masculine, feminine, and plural forms. The occupation and participant lexicon was reviewed by the adaptation team. Automated validation checks pair completeness, pronoun forms, entity inflection, annotation integrity, option-order balance, and exact reconstruction of clean sentences. Sixteen difficult pairs received explicit Ukrainian repairs.

Sentence-level human review is ongoing. This release is suitable for pipeline integration and preliminary evaluation; reported results should identify the dataset version and this review status.

Limitations

Ukrainian grammatical gender changes nouns, pronouns, adjectives, and verbs, so the variants are language-specific adaptations rather than literal substitutions. Masculine and feminine accuracy differences may reflect grammatical processing, translation choices, occupation associations, or their interaction. The forced-choice protocol is an added evaluation interface and should be distinguished from the original paper's resolver-based experiments.

License

The original WinoPron repository distributes its code and data under the GNU Affero General Public License v3.0. This Ukrainian adaptation is distributed under the same license. The original license text is included in LICENSE. The dataset was translated, grammatically adapted, and cleaned in 2026.

Original citation

@inproceedings{gautam-etal-2024-winopron,
  title = {{W}ino{P}ron: Revisiting {E}nglish {W}inogender Schemas for Consistency, Coverage, and Grammatical Case},
  author = {Gautam, Vagrant and Steuer, Julius and Bingert, Eileen and Johns, Ray and Lauscher, Anne and Klakow, Dietrich},
  booktitle = {Proceedings of the Seventh Workshop on Computational Models of Reference, Anaphora and Coreference},
  year = {2024},
  pages = {52--66},
  publisher = {Association for Computational Linguistics},
  doi = {10.18653/v1/2024.crac-1.6},
  url = {https://aclanthology.org/2024.crac-1.6/}
}

Ukrainian adaptation citation

@dataset{makovska_etal_2026_winopron_uk,
  title = {{W}ino{P}ron-{UK}},
  author = {Makovska, Viktoriia and Ostafiichuk, Oleksandra and Zakharchenko, Tetiana},
  year = {2026},
  publisher = {FairForget},
  url = {https://huggingface.co/datasets/FairForget/WinoPron-Uk},
  note = {Ukrainian translation, grammatical adaptation, and cleaning of WinoPron}
}