Generating Data to Mitigate Spurious Correlations in Natural Language Inference Datasets

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2660 - 2676 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Generating Data to Mitigate Spurious Correlations in Natural Language Inference Datasets Yuxiang Wu†∗ Matt Gardner ‡∗ Pontus Stenetorp † Pradeep Dasigi § † University College London yuxiang.wu,p.stenetorp@cs.ucl.ac.uk ‡ Microsoft Semantic Machines § Allen Institute for AI mattgardner@microsoft.com, pradeepd@allenai.org Abstract Natural language processing models often exploit spurious correlations between task- independent features and labels in datasets to perform well only within the distributions they are trained on, while not generalising to different task distributions. We propose to tackle this problem by generating a debi- ased version of a dataset, which can then be used to train a debiased, off-the-shelf model, by simply replacing its training data. Our approach consists of 1) a method for train- ing data generators to generate high-quality, label-consistent data samples; and 2) a filter- ing mechanism for removing data points that contribute to spurious correlations, measured in terms of z-statistics. We generate debiased versions of the SNLI and MNLI datasets,1 and we evaluate on a large suite of debiased, out- of-distribution, and adversarial test sets. Re- sults show that models trained on our debiased datasets generalise better than those trained on the original datasets in all settings. On the ma- jority of the datasets, our method outperforms or performs comparably to previous state-of- the-art debiasing strategies, and when com- bined with an orthogonal technique, product- of-experts, it improves further and outper- forms previous best results of SNLI-hard and MNLI-hard. 1 Introduction Natural Language Processing (NLP) datasets in- evitably contain biases that are unrelated to the tasks they are supposed to represent. These biases are usually artifacts of the annotation processes, task framing, or design decisions (Schwartz et al., 2017; Geva et al., 2019; Liu et al., 2021). Such biases often manifest as spurious correlations be- tween simple features of the data points and their ∗ Work done while at the Allen Institute for AI. 1All our code and the generated datasets are available at https://github.com/jimmycode/ gen-debiased-nli. Generator (Section 2 & Section 4.1) sample z-filter (Section 3 & Section 4.2) Figure 1: Overview of our dataset bias mitigation ap- proach. We minimise spurious correlations between la- bels (represented by the shapes of data points) and task- independent features (represented by their colours) with our proposed data generation pipeline. labels (Gardner et al., 2021). Trained models can exploit these spurious correlations to correctly pre- dict the labels of the data points within the same distributions as those they are trained on, but fail to generalise to other distributions within the same tasks. Consequently, the models risk modelling the datasets, but not the tasks (Gururangan et al., 2018; Poliak et al., 2018; McCoy et al., 2019; Schuster et al., 2019). We address this issue by adjusting existing dataset distributions to mitigate the correlations be- tween task-independent features and labels. First, we train data generators that generate high qual- ity data samples in …

Content truncated to 500 words