Debiasing Masks: A New Framework for Shortcut Mitigation in NLU

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 7607 - 7613 December 7-11, 2022 ©2022 Association for Computational Linguistics Debiasing Masks: A New Framework for Shortcut Mitigation in NLU Johannes Mario Meissner†, Saku Sugawara‡, Akiko Aizawa†‡ †The University of Tokyo, ‡National Institute of Informatics {meissner,saku,aizawa}@nii.ac.jp Abstract Debiasing language models from unwanted be- haviors in Natural Language Understanding tasks is a topic with rapidly increasing inter- est in the NLP community. Spurious statistical correlations in the data allow models to perform shortcuts and avoid uncovering more advanced and desirable linguistic features. A multitude of effective debiasing approaches has been pro- posed, but flexibility remains a major issue. For the most part, models must be retrained to find a new set of weights with debiased behavior. We propose a new debiasing method in which we identify debiased pruning masks that can be applied to a finetuned model. This enables the selective and conditional application of debias- ing behaviors. We assume that bias is caused by a certain subset of weights in the network; our method is, in essence, a mask search to iden- tify and remove biased weights. Our masks show equivalent or superior performance to the standard counterparts, while offering important benefits. Pruning masks can be stored with high efficiency in memory, and it becomes possible to switch among several debiasing behaviors (or revert back to the original biased model) at inference time. Finally, it opens the doors to further research on how biases are acquired by studying the generated masks. For example, we observed that the early layers and attention heads were pruned more aggressively, possibly hinting towards the location in which biases may be encoded. 1 Introduction The issue of spurious correlations in natural lan- guage understanding datasets has been extensively studied in recent years (Gururangan et al., 2018; McCoy et al., 2019; Gardner et al., 2021). In the MNLI dataset (Williams et al., 2018), negation words such as “not” are unintended hints for the contradiction label (Gururangan et al., 2018), while a high word overlap between the premise and the hypothesis often correlates with entailment (Mc- Coy et al., 2019). Transformer Layers Embedding Classifier Embedding Classifier Transformer Layers Embedding Classifier Biased Behavior Debiased Behavior 1 (e.g. lexical overlap) Debiased Behavior 2 (e.g. negation words) Transformer Layers 1. Freeze weights 2. Assign debiasing loss 3. Train scores (find mask) We mask individual weights in all modules, excluding embedding and classifier. Different debiasing targets lead to different masks. They can be applied to the same base model. Figure 1: We find masks that remove bias from a fine- tuned model by using frozen-weight movement pruning. A common way to prevent the acquisition of bi- ases (in this paper defined as unintended shortcut behaviors) is to adopt a debiasing loss function, such as product-of-experts (Hinton, 2002), to dis- courage learning shortcuts through the use of the annotated level of bias in each sample. Because manual bias annotation is difficult and expensive, the predictions of a biased …

Content truncated to 500 words