Fair Hate Speech Detection through Evaluation of Social Group Counterfactuals
Aida Mostafazadeh Davani1, Ali Omrani1, Brendan Kennedy1, Mohammad Atari2, Xiang Ren1, Morteza Dehghani12
Abstract
Approaches for mitigating bias in supervised models are designed to reduce models’ dependence on specific sensitive features of the input data, e.g., mentioned social groups. However, in the case of hate speech detection, it is not always desirable to equalize the effects of social groups because of their essential role in distinguishing outgroup-derogatory hate, such that particular types of hateful rhetoric carry the intended meaning only when contextualized around certain social group tokens. Counterfactual token fairness for a mentioned social group evaluates the model’s predictions as to whether they are the same for (a) the actual sentence and (b) a counterfactual instance, which is generated by changing the mentioned social group in the sentence. Our approach assures robust model predictions for counterfactuals that imply similar meaning as the actual sentence. To quantify the similarity of a sentence and its counterfactual, we compare their likelihood score calculated by generative language models. By equalizing model behaviors on each sentence and its counterfactuals, we mitigate bias in the proposed model while preserving the overall classification performance.
中文速览
仇恨言论检测模型常因训练数据中某些社会群体词汇（Social Group Tokens，SGTs，如"穆斯林""犹太人"）出现频率过高而产生误判偏差，现有反偏差方法通常要求模型对所有"反事实替换"（即把句中的群体词换成另一个群体词）给出相同预测，但这在仇恨言论场景中并不合理——许多仇恨表述的含义本身就与特定群体的刻板印象紧密绑定，强行替换会改变句子语义。针对这一问题，本文提出用预训练语言模型 GPT-2 计算句子对数似然，来判断一次群体词替换是否保留了原句语义：若替换后语义差异过大（"非对称反事实"），则在训练时将其排除，只对语义相近的替换（"对称反事实"）施加逻辑值配对（Counterfactual Logit Pairing）约束，使模型在这些真正可比的情况下保持一致预测。在 Stormfront 和 Gab Hate Corpus 两个数据集上的实验表明，该方法在保持仇恨言论分类性能的同时，比现有去偏方法更有效地提升了反事实公平性，为需要区分群体特异性语义的 NLP 公平性研究提供了一个更精细的处理框架。
原文 arXiv:2010.12779;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2010.12779v1