Increasing Robustness to Spurious Correlations using Forgettable Examples

Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics, pages 3319–3332 April 19 - 23, 2021. ©2021 Association for Computational Linguistics 3319 Increasing Robustness to Spurious Correlations using Forgettable Examples Yadollah Yaghoobzadeh1 Soroush Mehri2 Remi Tachet des Combes2 Timothy J. Hazen1 Alessandro Sordoni2 1Microsoft Turing, Montr´eal 2Microsoft Research, Montr´eal {yayaghoo,alsordon}@microsoft.com Abstract Neural NLP models tend to rely on spuri- ous correlations between labels and input fea- tures to perform their tasks. Minority exam- ples, i.e., examples that contradict the spuri- ous correlations present in the majority of data points, have been shown to increase the out-of- distribution generalization of pre-trained lan- guage models. In this paper, we first propose using example forgetting to find minority ex- amples without prior knowledge of the spuri- ous correlations present in the dataset. Forget- table examples are instances either learned and then forgotten during training or never learned. We empirically show how these examples are related to minorities in our training sets. Then, we introduce a new approach to robustify mod- els by fine-tuning our models twice, first on the full training data and second on the minori- ties only. We obtain substantial improvements in out-of-distribution generalization when ap- plying our approach to the MNLI, QQP, and FEVER datasets. 1 Introduction Despite the impressive performance of current NLP models, these models often exploit spurious corre- lations: they tend to capture prediction correlations that hold for most examples but do not hold in gen- eral. For instance, in natural language inference (NLI) datasets, word-overlap between hypothesis and premise is highly correlated with the entail- ment label (McCoy et al., 2019; Zhang et al., 2019). Therefore, these models are brittle when tested on examples that cannot be solved by recurring to these correlations, limiting their application in real- world scenarios. Out-of-distribution or challenging sets are benchmarks carefully designed to break systems that rely on such correlations. The paradigm of fine-tuning pre-trained lan- guage models (PLM) has pushed the state-of-the- art in a large variety of tasks involving natural lan- guage understanding (NLU) (Devlin et al., 2019; Wang et al., 2019). This is achieved by self- supervised learning from an enormous amount of text. PLMs also show increased robustness on chal- lenging datasets (Hendrycks et al., 2019). This increase is attributed to an empirical finding that PLMs perform better on minority examples present in the training data (Tu et al., 2020). These minor- ity examples violate the spurious correlations and therefore likely support the examples in challeng- ing datasets. Tu et al. (2020) find minority examples by man- ually dividing the training data into two groups, according to the known spurious correlations (e.g., word-overlap in NLI). They present an analysis of the robustness of PLMs and its connection to mi- nority examples. In this work, we first introduce a systematic way to find minority examples that does not need prior knowledge of spurious corre- lations, a big limitation of the earlier work. We then present a simple approach that increases the …

Content truncated to 500 words