A few large, homogenous, pre-trained models undergird many machine learning systems — and often, these models contain harmful stereotypes learned from the internet. We investigate the bias transfer hypothesis: the theory that social biases (such as stereotypes) internalized by large language models during pre-training transfer into harmful task-specific behavior after fine-tuning. For two classification tasks, we find that reducing intrinsic bias with controlled interventions before fine-tuning does little to mitigate the classifier’s discriminatory behavior after fine-tuning. Regression analysis suggests that downstream disparities are better explained by biases in the fine-tuning dataset. Still, pre-training plays a role: simple alterations to co-occurrence rates in the fine-tuning dataset are ineffective when the model has been pre-trained. Our results encourage practitioners to focus more on dataset quality and context-specific harms.

Transformer-based pre-trained models are known to encode societal biases not only in their contextual representations, but also in downstream predictions when fine-tuned on task-specific data. We present D-Bias, an approach that selectively eliminates stereotypical associations (e.g, co-occurrence statistics) at fine-tuning, such that the model doesn’t learn to excessively rely on those signals.D-Bias attenuates biases from both identity words and frequently co-occurring proxies, which we select using pointwise mutual information. We apply D-Bias to a) occupation classification, and b) toxicity classification and find that our approach substantially reduces downstream biases (e.g. by > 60% in toxicity classification, for identities that are most frequently flagged as toxic on online platforms).In addition, we show that D-Bias dramatically improves upon scrubbing, i.e., removing only the identity words in question. We also demonstrate that D-Bias easily extends to multiple identities, and achieves competitive performance with two recently proposed debiasing approaches: R-LACE and INLP.

We present a method for adversarial input generation against black box models for reading comprehension based question answering. Our approach is composed of two steps. First, we approximate a victim black box model via model extraction. Second, we use our own white box method to generate input perturbations that cause the approximate model to fail. These perturbed inputs are used against the victim. In experiments we find that our method improves on the efficacy of the ADDANY—a white box attack—performed on the approximate model by 25% F1, and the ADDSENT attack—a black box attack—by 11% F1.

String similarity models are vital for record linkage, entity resolution, and search. In this work, we present STANCE–a learned model for computing the similarity of two strings. Our approach encodes the characters of each string, aligns the encodings using Sinkhorn Iteration (alignment is posed as an instance of optimal transport) and scores the alignment with a convolutional neural network. We evaluate STANCE’s ability to detect whether two strings can refer to the same entity–a task we term alias detection. We construct five new alias detection datasets (and make them publicly available). We show that STANCE (or one of its variants) outperforms both state-of-the-art and classic, parameter-free similarity models on four of the five datasets. We also demonstrate STANCE’s ability to improve downstream tasks by applying it to an instance of cross-document coreference and show that it leads to a 2.8 point improvement in Bˆ3 F1 over the previous state-of-the-art approach.