Gender Bias in Neural Natural Language Processing
Kaiji Lu Carnegie Mellon University Moffiet Field, CA 94035、Piotr Mardziel Carnegie Mellon University Moffiet Field, CA 94035、Fangjing Wu Carnegie Mellon University Moffiet Field, CA 94035、Preetam Amancharla Carnegie Mellon University Moffiet Field, CA 94035、Anupam Datta Carnegie Mellon University Moffiet Field, CA 94035
Abstract
We examine whether neural natural language processing (NLP) systems reflect historical biases in training data. We define a general benchmark to quantify gender bias in a variety of neural NLP tasks. Our empirical evaluation with state-of-the-art neural coreference resolution and textbook RNN-based language models trained on benchmark data sets finds significant gender bias in how models view occupations. We then mitigate bias with counterfactual data augmentation (CDA): a generic methodology for corpus augmentation via causal interventions that breaks associations between gendered and gender-neutral words. We empirically show that CDA effectively decreases gender bias while preserving accuracy. We also explore the space of mitigation strategies with CDA, a prior approach to word embedding debiasing (WED), and their compositions. We show that CDA outperforms WED, drastically so when word embeddings are trained. For pre-trained embeddings, the two methods can be effectively composed. We also find that as training proceeds on the original data set with gradient descent the gender bias grows as the loss reduces, indicating that the optimization encourages bias; CDA mitigates this beha
中文速览
神经网络自然语言处理模型会从训练数据中"学会"性别偏见——比如把"医生"更多地与"他"而非"她"关联起来，但如何系统地衡量和消除这种偏见此前并无通用方案。研究者提出了一套通用基准，通过对比仅改变性别词的配对句子来量化模型在指代消解（coreference resolution）和语言模型（language model）任务中的性别偏见，实验证实当前最先进的神经网络模型在职业词汇上存在显著偏见，且随着训练推进、损失下降，偏见反而会加剧。为此，他们提出"反事实数据增强"（counterfactual data augmentation，CDA）方法：在每条训练样本旁加入一条性别词互换的镜像样本，迫使模型无法依赖性别线索做出预测。实验表明，CDA在保持模型准确率的同时能有效降低偏见，效果优于此前流行的词向量去偏（word embedding debiasing，WED）方法，两者在使用预训练词向量时还可叠加使用，为构建更公平的NLP系统提供了切实可行的工具。
原文 arXiv:1807.11714;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1807.11714v2