crows_swapped / README.md
borabozdogan's picture
Add license, attribution and modification notice
b50204c verified
|
Raw History Blame Contribute Delete
3.49 kB
metadata
license: cc-by-sa-4.0
language:
  - en
pretty_name: CrowS-Pairs (Synonym-Swapped)
tags:
  - bias
  - stereotypes
  - fairness
  - robustness
  - synonym-substitution
dataset_info:
  features:
    - name: id
      dtype: int64
    - name: sent_more
      dtype: string
    - name: sent_less
      dtype: string
    - name: bias_type
      dtype:
        class_label:
          names:
            '0': race-color
            '1': socioeconomic
            '2': gender
            '3': disability
            '4': nationality
            '5': sexual-orientation
            '6': physical-appearance
            '7': religion
            '8': age
    - name: stereo_antistereo
      dtype:
        class_label:
          names:
            '0': stereo
            '1': antistereo
    - name: annotations
      dtype: string
    - name: anon_writer
      dtype: string
    - name: anon_annotators
      dtype: string
  splits:
    - name: test
      num_bytes: 390419
      num_examples: 1508
  download_size: 201817
  dataset_size: 390419
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

CrowS-Pairs (Synonym-Swapped)

This is a modified version of CrowS-Pairs (Nangia et al., 2020), redistributed under the original CC BY-SA 4.0 license.

What changed

sent_more and sent_less were rewritten with the same synonym substitutions, so each pair still differs only in its demographic terms. Words that differ between the two sentences were excluded from swapping. IDs, bias types, labels and annotations are unchanged.

Synonyms were proposed automatically by an LLM (gpt-4o-mini), which was asked to suggest close synonyms while skipping stopwords, proper nouns, numbers and words central to the sentence's topic. Every proposed substitution was applied (swap probability 1.0). Because the substitutions are automatic, a few may slightly shift meaning.

The dataset is intended for testing whether bias-benchmark results hold up under surface-level rewording.

License

CC BY-SA 4.0, inherited from the original dataset; the full text is in LICENSE. Under the ShareAlike term, adaptations of this dataset must also be distributed under CC BY-SA 4.0.

The original CrowS-Pairs was created using prompts from the ROCStories corpora and the fiction section of MNLI. Its repository states the CC BY-SA 4.0 license in its README; the LICENSE file here is the official CC BY-SA 4.0 legal code.

Content warning

Like the original, this dataset contains stereotypes and offensive content about social groups. It is meant for measuring bias, not for training models to produce such content.

Citation

If you use this dataset, please cite the original CrowS-Pairs paper:

@inproceedings{nangia-etal-2020-crows,
    title = "{C}row{S}-Pairs: A Challenge Dataset for Measuring Social Biases in Masked Language Models",
    author = "Nangia, Nikita  and
      Vania, Clara  and
      Bhalerao, Rasika  and
      Bowman, Samuel R.",
    editor = "Webber, Bonnie  and
      Cohn, Trevor  and
      He, Yulan  and
      Liu, Yang",
    booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing ({EMNLP})",
    month = nov,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2020.emnlp-main.154/",
    doi = "10.18653/v1/2020.emnlp-main.154",
    pages = "1953--1967"
}