Controlling Learned Effects to Reduce Spurious Correlations in Text Classifiers

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2271–2287 July 9-14, 2023 ©2023 Association for Computational Linguistics Controlling Learned Effects to Reduce Spurious Correlations in Text Classifiers Parikshit Bansal Microsoft Research, India parikshitb52@gmail.com Amit Sharma Microsoft Research, India amshar@microsoft.com Abstract To address the problem of NLP classifiers learn- ing spurious correlations between training fea- tures and target labels, a common approach is to make the model’s predictions invariant to these features. However, this can be counter- productive when the features have a non-zero causal effect on the target label and thus are im- portant for prediction. Therefore, using meth- ods from the causal inference literature, we propose an algorithm to regularize the learnt effect of the features on the model’s prediction to the estimated effect of feature on label. This results in an automated augmentation method that leverages the estimated effect of a feature to appropriately change the labels for new aug- mented inputs. On toxicity and IMDB review datasets, the proposed algorithm minimises spu- rious correlations and improves the minority group (i.e., samples breaking spurious correla- tions) accuracy, while also improving the total accuracy compared to standard training. 1 1 Introduction While classifiers trained on pre-trained NLP mod- els achieve state-of-the-art accuracy on various tasks, they have been shown to learn spurious cor- relations between input features and the label (Du et al., 2022). Such learned correlations impact ac- curacy on out-of-distribution samples and in the case of sensitive spurious features, lead to unfair predictions (Sun et al., 2019; Ribeiro et al., 2020). Learned spurious correlations can be over features that are either irrelevant (e.g., tense, gender for pro- fession classification) or relevant (e.g., emoticons for sentiment classification, negation words for con- tradiction). In both cases, the classifier overweighs their importance compared to other features. For removing spurious correlations, a common principle underlying past work is to make a model’s prediction invariant to the features that exhibit the 1Code: https://github.com/pbansal5/ feature-effect-augmentation Figure 1: Example from IMDB reviews dataset show- ing the spurious token “8/10” and its importance for prediction on some inputs. Parts highlighted in yellow are ambiguous in sentiment, in green are (supposedly) positive in sentiment and red are negative. correlation. This can be done by data augmentation (Kaushik et al., 2019), latent space removal (Ravfo- gel et al., 2020), subsampling (Sagawa et al., 2019, 2020), or sample reweighing (Mahabadi et al., 2019; Orgad and Belinkov, 2022). In many cases, however, the correlated features may be important for the task and their complete removal can cause a degradation in task performance. For instance, for spurious correlation over negation tokens (e.g., “not”) or lexical overlap in MNLI natural language inference tasks, Williams et al. (2017); Joshi et al. (2022) show that correlated features are necessary for prediction and their removal can hurt accuracy. As another example, consider the IMDB review dataset (Maas et al., 2011) where the task is clas- sify the sentiment of a given review as positive or …

Content truncated to 500 words