wbaek / kobbq

Official code and dataset repository of KoBBQ (TACL 2024)

KoBBQ

This is the official repository of KoBBQ: Korean Bias Benchmark for Question Answering (TACL 2024).

Webpage

• https://jinjh0123.github.io/KoBBQ

Datasets

• Our KoBBQ datasets and survey results can be found in KoBBQ/data & Huggingface Datasets.

Examples

Statistics

Category	# of Templates	# of Samples
Age	21	3,608
Disability Status	20	2,160
Gender Identity	25	768
Physical Appearance	20	4,040
Race/Ethnicity/Nationality	43	51,856
Religion	20	688
Socio-Economic Status	27	6,928
Sexual Orientation	12	552
Domestic Area of Origin	22	800
Family Structure	23	1,096
Political Orientation	11	312
Education Background	24	3,240
Total	268	76,048

How to Evaluate

• Our evaluation codes and prompts can be found in KoBBQ/evaluation.

With Evaluation Set

1. Put model outputs to prediction column in KoBBQ/data/KoBBQ_test_samples.tsv and save the file as KoBBQ/evaluation/outputs/KoBBQ_test/KoBBQ_test_evaluation_1_{$MODEL}.tsv.
   • The model outputs should be one of the choices (as in choices column). Otherwise, they will be regarded as out-of-choice answers.
2. Run KoBBQ/evaluation/5_evaluation.py with test option.

   cd evaluation
   python3 5_evaluation.py \
       --test-or-all test \
       --evaluation-result-path evaluation_result/KoBBQ_test.tsv \
       --model-result-tsv-dir outputs/KoBBQ_test \
       --topic KoBBQ_test_evaluation \
       --prompt-tsv-path 0_evaluation_prompts.tsv \
       --prompt-id 1 \
       --models $MODEL

With Entire Set

1. Put model outputs to prediction column in KoBBQ/data/KoBBQ_all_samples.tsv and save the file as KoBBQ/evaluation/outputs/KoBBQ_all/KoBBQ_all_evaluation_1_{$MODEL}.tsv.
   • The model outputs should be one of the choices (as in choices column). Otherwise, they will be regarded as out-of-choice answers.
2. Run KoBBQ/evaluation/5_evaluation.py with all option.

   cd evaluation
   python3 5_evaluation.py \
       --test-or-all all \
       --evaluation-result-path evaluation_result/KoBBQ_all.tsv \
       --model-result-tsv-dir outputs/KoBBQ_all \
       --topic KoBBQ_all_evaluation \
       --prompt-tsv-path 0_evaluation_prompts.tsv \
       --prompt-id 1 \
       --models $MODEL

Ethical Considerations

• We do not condone any malicious use of our dataset. It must not be used as training data to automatically generate and publish biased languages targeting specific groups. We strongly encourage researchers and practitioners to utilize this dataset in beneficial ways, such as mitigating bias in language models.

BibTex

@article{10.1162/tacl_a_00661,
    author = {Jin, Jiho and Kim, Jiseon and Lee, Nayeon and Yoo, Haneul and Oh, Alice and Lee, Hwaran},
    title = "{KoBBQ: Korean Bias Benchmark for Question Answering}",
    journal = {Transactions of the Association for Computational Linguistics},
    volume = {12},
    pages = {507-524},
    year = {2024},
    month = {05},
    issn = {2307-387X},
    doi = {10.1162/tacl_a_00661},
    url = {https://doi.org/10.1162/tacl_a_00661},
    eprint = {https://direct.mit.edu/tacl/article-pdf/doi/10.1162/tacl_a_00661/2369542/tacl_a_00661.pdf}
}