winoBias-eval
Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods — Zhao et al. (2018) (arXiv:1804.06876, 2018)
What this evaluates
Evaluates gender bias in coreference resolution systems by measuring performance disparity between pro-stereotypical and anti-stereotypical sentences. It probes whether models rely on gender stereotypes when resolving coreferences in challenging, Winograd-style contexts.
Datasets
- WinoBias — total ?; splits: dev (-1), test (-1)
Metrics
F1(primary) — range: [0, 100]- Standard coreference F1 score computed using scorer v8.01 (Pradhan et al., 2014).
|Diff|— range: [0, 100]- Absolute difference between F1 scores on pro-stereotyped and anti-stereotyped subsets.
Input / output format
Input: Winograd-style sentences containing occupation-based entities with gendered pronouns, presented in pro-stereotypical and anti-stereotypical conditions.
Output: Set of coreference links (pairs of entity mentions) for each input sentence.
Scoring recipe
pred_links = model.predict(sentence)
gold_links = get_gold_coref(sentence)
f1 = scorer_v801.compute_f1(pred_links, gold_links)
f1_pro = compute_f1(pro_stereotypical_sentences)
f1_anti = compute_f1(anti_stereotypical_sentences)
diff = abs(f1_pro - f1_anti)
Common pitfalls
- Using a coreference scorer other than v8.01 (Pradhan et al., 2014), which yields different F1 values.
- Ignoring statistical significance testing for |Diff|; the paper requires an approximate randomized test (p < .05) to flag significant bias.
- Evaluating on standard coreference benchmarks alone, which masks bias that only appears in WinoBias's adversarial pro/anti-stereotypical split.
Evidence (verbatim from paper)
Table 2: F1 on OntoNotes and WinoBias development set. WinoBias results are split between Type-1 and Type-2 and in pro/anti-stereotypical conditions.
Citation
@misc{zhao2018genderbias,
title={Gender Bias in Coreference Resolution: Evaluation and Debiasing Methods},
author={Zhao et al. (2018)},
year={2018},
note={arXiv:1804.06876}
}
- arXiv: 1804.06876