On Measuring and Mitigating Biased Inferences of Word Embeddings
Sunipa Dev, Tao Li, Jeff M. Phillips, Vivek Srikumar School of Computing University of Utah Salt Lake City, Utah, USA {sunipad, tli, jeffp,
Abstract
Word embeddings carry stereotypical connotations from the text they are trained on, which can lead to invalid inferences in downstream models that rely on them. We use this observation to design a mechanism for measuring stereotypes using the task of natural language inference. We demonstrate a reduction in invalid inferences via bias mitigation strategies on static word embeddings (GloVe). Further, we show that for gender bias, these techniques extend to contextualized embeddings when applied selectively only to the static components of contextualized embeddings (ELMo, BERT).
中文速览
词向量（word embedding）中藏着对性别、国籍、宗教等群体的刻板印象，会让下游模型做出错误推断，但现有的偏见检测方法大多只看词向量之间的距离，既与实际使用方式脱节，也难以处理随上下文变化的动态向量。作者提出用自然语言推理（NLI）任务来衡量偏见：构造数百万个"前提与假设理应中性"的句对，让模型去判断蕴含关系，若模型频繁给出非中性预测，就说明词向量携带了偏见信息。实验发现GloVe、ELMo和BERT都存在显著的性别、国籍与宗教偏见；进一步测试去偏方法后发现，对静态词向量（GloVe）直接做投影去偏效果明显，而对动态向量（ELMo、BERT）则需仅针对其静态底层部分（ELMo的第一层、BERT的子词嵌入）施加去偏才能有效降低性别偏见，但对国籍和宗教偏见的去除仍收效有限。这项工作为偏见的测量与缓解提供了一套更贴近实际下游任务的评估框架，并首次系统展示了主流词向量中存在国籍与宗教偏见。
原文 arXiv:1908.09369;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1908.09369v3