Decorrelate Irrelevant, Purify Relevant: Overcome Textual Spurious Correlations from a Feature Perspective

Proceedings of the 29th International Conference on Computational Linguistics, pages 2278–2287 October 12–17, 2022. 2278 Decorrelate Irrelevant, Purify Relevant: Overcome Textual Spurious Correlations from a Feature Perspective Shihan Dou1∗, Rui Zheng1∗, Ting Wu1, Songyang Gao1, Junjie Shan3, Qi Zhang12, Yueming Wu4, Xuanjing Huang1† 1 School of Computer Science, Fudan University, Shanghai, China 2 Shanghai Key Laboratory of Intelligent Information Processing, Fudan University 3 KTH Royal Institute of Technology, Stockholm, Sweden 4 Nanyang Technological University, Singapore {shdou21,tingwu21,gaosy21}@m.fudan.edu.cn {rzheng20,qz,xjhuang}@fudan.edu.cn Abstract Natural language understanding (NLU) mod- els tend to rely on spurious correlations (i.e., dataset bias) to achieve high performance on in-distribution datasets but poor performance on out-of-distribution ones. Most of the ex- isting debiasing methods often identify and weaken these samples with biased features (i.e., superficial surface features that cause such spu- rious correlations). However, down-weighting these samples obstructs the model in learning from the non-biased parts of these samples. To tackle this challenge, in this paper, we pro- pose to eliminate spurious correlations in a fine-grained manner from a feature space per- spective. Specifically, we introduce Random Fourier Features and weighted re-sampling to decorrelate the dependencies between features to mitigate spurious correlations. After obtain- ing decorrelated features, we further design a mutual-information-based method to purify them, which forces the model to learn features that are more relevant to tasks. Extensive exper- iments on two well-studied NLU tasks demon- strate that our method is superior to other com- parative approaches. 1 Introduction Recently, researchers have found that the main rea- son why large-scale pre-trained language models perform well on NLU tasks is that they rely on spurious correlations, rather than capturing the lan- guage understanding for the intended task (Bender and Koller, 2020). These spurious correlations are also denoted as dataset bias in previous work (He et al., 2019; Clark et al., 2019): prediction rules that work for training examples but do not hold in general. In reality, a variety of spurious correlations appear in widely-used NLU benchmark datasets. ∗ Equal contribution. † Corresponding author. For example, in natural language inference (NLI) tasks, McCoy et al. (2019) observe that models on the MNLI dataset (Williams et al., 2018) rely heav- ily on the features of word overlap to predict the entailment label blindly. Consequently, these mod- els perform poorly on out-of-distribution (OOD) datasets where such correlations no longer hold (Nie et al., 2019). To mitigate these spurious correlations, some ex- isting debiasing works (Clark et al., 2019; He et al., 2019) prefer to train a bias model with known spuri- ous correlations as prior knowledge to identify the samples without biased features. This trained bias model is used in the later stage to force the main model to learn from these samples. For better trans- ferability, Utama et al. (2020b); Sanh et al. (2020) relax this basic assumption that spurious correla- tion is apriori by using a small part of the training dataset in the training phase of bias model. How- ever, these methods are not end-to-end and their training procedures are …

Content truncated to 500 words