Is Your Classifier Actually Biased? Measuring Fairness under Uncertainty with Bernstein Bounds
Is Your Classifier Actually Biased? Measuring Fairness under Uncertainty with Bernstein Bounds
复制标题
你的分类器实际上有偏见吗？
DOI：
10.18653/v1/2020.acl-main.262
复制
发表时间：
2020
影响因子：
6
通讯作者：
Kawin Ethayarajh
中科院分区：
文献类型：
--
作者：
Kawin Ethayarajh
大多数NLP数据集都没有使用受保护的属性（如性别）进行注释，因此很难使用标准的公平性度量（例如，机会均等）。但是，手动注释具有受保护属性的大型数据集是缓慢且昂贵的。我们能不能不标注所有的例子，而是标注其中的一个子集，并使用该样本来估计偏差？虽然有可能这样做，但注释样本越小，我们就越不确定估计值是否接近真实偏倚。在这项工作中，我们建议使用伯恩斯坦界来表示这种不确定性的偏差估计的置信区间。我们提供的经验证据表明，以这种方式得出的95%置信区间始终是真实偏倚的界限。在量化这种不确定性时，我们称之为伯恩斯坦有界不公平的方法有助于防止分类器在没有足够的证据做出任何声明时被认为是有偏见或无偏见的。我们的研究结果表明，目前用于测量特定偏差的数据集太小，无法最终确定偏差，除非在最严重的情况下。例如，考虑一个共指消解系统，它在性别刻板印象句子上的准确率要高出5%-要声称它有95%的置信度，我们需要一个比WinoBias大3.8倍的偏差特定数据集，WinoBias是最大的可用数据集。
Most NLP datasets are not annotated with protected attributes such as gender, making it difficult to measure classification bias using standard measures of fairness (e.g., equal opportunity). However, manually annotating a large dataset with a protected attribute is slow and expensive. Instead of annotating all the examples, can we annotate a subset of them and use that sample to estimate the bias? While it is possible to do so, the smaller this annotated sample is, the less certain we are that the estimate is close to the true bias. In this work, we propose using Bernstein bounds to represent this uncertainty about the bias estimate as a confidence interval. We provide empirical evidence that a 95% confidence interval derived this way consistently bounds the true bias. In quantifying this uncertainty, our method, which we call Bernstein-bounded unfairness, helps prevent classifiers from being deemed biased or unbiased when there is insufficient evidence to make either claim. Our findings suggest that the datasets currently used to measure specific biases are too small to conclusively identify bias except in the most egregious cases. For example, consider a co-reference resolution system that is 5% more accurate on gender-stereotypical sentences – to claim it is biased with 95% confidence, we need a bias-specific dataset that is 3.8 times larger than WinoBias, the largest available.
DOI：
10.1073/pnas.1720347115
发表时间：
2018-04-17
影响因子：
11.1
作者：
Garg, Nikhil;Schiebinger, Londa;Zou, James
通讯作者：
Zou, James
黑人之于罪犯就像白人之于警察：检测和消除词嵌入中的多类偏差
DOI：
--
发表时间：
2019
期刊：
2019 Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL
影响因子：
--
作者：
Manzini, Thomas;Lim, Yao Chong;Tsvetkov, Yulia;Black, Alan W
通讯作者：
Black, Alan W
DOI：
10.18653/v1/n18-2003
发表时间：
2018-04
期刊：
ArXiv
影响因子：
--
作者：
Jieyu Zhao;Tianlu Wang;Mark Yatskar;Vicente Ordonez;Kai-Wei Chang
通讯作者：
Jieyu Zhao;Tianlu Wang;Mark Yatskar;Vicente Ordonez;Kai-Wei Chang