Bias mitigation approaches reduce models dependence on sensitive features of data, such as social group tokens (SGTs), resulting in equal predictions across the sensitive features. In hate speech detection, however, equalizing model predictions may i
gnore important differences among targeted social groups, as hate speech can contain stereotypical language specific to each SGT. Here, to take the specific language about each SGT into account, we rely on counterfactual fairness and equalize predictions among counterfactuals, generated by changing the SGTs. Our method evaluates the similarity in sentence likelihoods (via pre-trained language models) among counterfactuals, to treat SGTs equally only within interchangeable contexts. By applying logit pairing to equalize outcomes on the restricted set of counterfactuals for each instance, we improve fairness metrics while preserving model performance on hate speech detection.
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 models 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.
