Improving the robustness of NLI models with minimax training

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 14322–14339 July 9-14, 2023 ©2023 Association for Computational Linguistics Improving the robustness of NLI models with minimax training Michalis Korakakis University of Cambridge mk2008@cam.ac.uk Andreas Vlachos University of Cambridge av308@cam.ac.uk Abstract Natural language inference (NLI) models are susceptible to learning shortcuts, i.e. decision rules that spuriously correlate with the label. As a result, they achieve high in-distribution performance, but fail to generalize to out-of- distribution samples where such correlations do not hold. In this paper, we present a training method to reduce the reliance of NLI models on shortcuts and improve their out-of-distribution performance without assuming prior knowl- edge of the shortcuts being targeted. To this end, we propose a minimax objective between a learner model being trained for the NLI task, and an auxiliary model aiming to maximize the learner’s loss by up-weighting examples from regions of the input space where the learner incurs high losses. This process incentivizes the learner to focus on under-represented “hard” examples with patterns that contradict the short- cuts learned from the prevailing “easy” ex- amples. Experimental results on three NLI datasets demonstrate that our method consis- tently outperforms other robustness enhancing techniques on out-of-distribution adversarial test sets, while maintaining high in-distribution accuracy. 1 Introduction Natural language inference (NLI)1 models have achieved state-of-the-art results on many bench- marks (Wang et al., 2019). However, recent work has demonstrated that their success is partly due to learning and using shortcuts (Gururangan et al., 2018; Poliak et al., 2018; Geirhos et al., 2020), i.e. spurious correlations between input attributes and labels introduced during dataset creation.2 For example, high word-overlap between the premise and the hypothesis in the MNLI (Williams et al., 1Also known as textual entailment (Dagan et al., 2005). 2In this paper, we use the terminology introduced by Geirhos et al. (2020). Other works also refer to shortcuts as “biases” and/or “heuristics.” Learner min L Training Data Auxiliary max L Example Weights Predictions Figure 1: Illustration of the proposed minimax train- ing objective. The learner optimizes for the NLI task, whereas the auxiliary tries to maximize the learner’s loss by generating an example weight distribution that encourages the learner to focus on “hard” examples with patterns that contradict the shortcuts. At inference time the learner can make predictions without the auxiliary. 2018) dataset is strongly correlated with the entail- ment label (McCoy et al., 2019). Consequently, NLI models that exploit shortcuts perform well on in-distribution samples, but are brittle when tested on out-of-distribution adversarial test sets that ex- plicitly target such phenomena (Naik et al., 2018; Glockner et al., 2018). Thus, numerous methods have been proposed to prevent models from learning shortcuts present in NLI datasets (Belinkov et al., 2019; Schuster et al., 2019; Zhou and Bansal, 2020; Stacey et al., 2020; Du et al., 2021; Modarressi et al., 2023, inter alia). Most approaches typically assume access to an auxiliary model designed to rely on shortcuts for predictions. The …

Content truncated to 500 words