Bias-measuring datasets play a critical role in detecting biased behavior of language models and in evaluating progress of bias mitigation methods. In this work, we focus on evaluating gender bias through coreference resolution, where previous datasets are either hand-crafted or fail to reliably measure an explicitly defined bias. To overcome these shortcomings, we propose a novel method to collect diverse, natural, and minimally distant text pairs via counterfactual generation, and construct Counter-GAP, an annotated dataset consisting of 4008 instances grouped into 1002 quadruples. We further identify a bias cancellation problem in previous group-level metrics on Counter-GAP, and propose to use the difference between inconsistency across genders and within genders to measure bias at a quadruple level. Our results show that four pre-trained language models are significantly more inconsistent across different gender groups than within each group, and that a name-based counterfactual data augmentation method is more effective to mitigate such bias than an anonymization-based method.

The increasingly large size of modern pre-trained language models not only makes them inherit more human-like biases from the training corpora, but also makes it computationally expensive to mitigate such biases. In this paper, we investigate recent parameter-efficient methods in combination with counterfactual data augmentation (CDA) for bias mitigation. We conduct extensive experiments with prefix tuning, prompt tuning, and adapter tuning on different language models and bias types to evaluate their debiasing performance and abilities to preserve the internal knowledge of a pre-trained model. We find that the parameter-efficient methods (i) are effective in mitigating gender bias, where adapter tuning is consistently the most effective one and prompt tuning is more suitable for GPT-2 than BERT, (ii) areless effective when it comes to racial and religious bias, which may be attributed to the limitations of CDA, and (iii) can perform similarly to or sometimes better than full fine-tuning with improved time and memory efficiency, as well as maintain the internal knowledge in BERT and GPT-2, evaluated via fact retrieval and downstream fine-tuning.