Kernel-Whitening: Overcome Dataset Bias with Isotropic Sentence Embedding

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4112 - 4122 December 7-11, 2022 ©2022 Association for Computational Linguistics Kernel-Whitening: Overcome Dataset Bias with Isotropic Sentence Embedding Songyang Gao1, Shihan Dou1, Qi Zhang12∗, Xuanjing Huang12 1 School of Computer Science, Fudan University, Shanghai, China 2 Shanghai Key Laboratory of Intelligent Information Processing, Shanghai, China {gaosy21, shdou21}@m.fudan.edu.cn Abstract Dataset bias has attracted increasing attention recently for its detrimental effect on the gen- eralization ability of fine-tuned models. The current mainstream solution is designing an ad- ditional shallow model to pre-identify biased instances. However, such two-stage methods scale up the computational complexity of train- ing process and obstruct valid feature informa- tion while mitigating bias. To address this is- sue, we utilize the representation normalization method which aims at disentangling the corre- lations between features of encoded sentences. We find it also promising in eliminating the bias problem by providing isotropic data distri- bution. We further propose Kernel-Whitening, a Nystr¨om kernel approximation method to achieve more thorough debiasing on nonlin- ear spurious correlations. Our framework is end-to-end with similar time consumption to fine-tuning. Experiments show that Kernel- Whitening significantly improves the perfor- mance of BERT on out-of-distribution datasets while maintaining in-distribution accuracy. 1 Introduction Despite remarkable performance on NLP tasks, pre- trained language models, like BERT, suffer sharp performance degradation in out-of-distribution (OOD) settings (McCoy et al., 2019). The above de- fect roots in the excessive reliance on spurious cor- relations, which is widely found in crowdsourcing- built datasets (Gururangan et al., 2018). These phe- nomena are donated as dataset bias problem (He et al., 2019). A line of works attempts to tackle this problem by down-weighting bias training exam- ples to discourage the main model from adopting recognized biases, including example reweighting (Schuster et al., 2019), confidence regularization (Utama et al., 2020a), or model ensembling (Clark et al., 2019). ∗ Corresponding author. : OOD Samples : Decision Boundaries : ID Samples Isotropic Latent Space Sentence Embedding Space Transform Inverse Figure 1: Illustration of Kernel-Whitening. The vertical and horizontal axes represent the valid and invalid fea- tures, respectively. Uneven sample distribution induces a bias decision boundary, resulting in errors on out-of- distribution data. The normalization method maps the data to isotropic latent space, where the new boundary is uncorrelated to redundant features, providing better generalization. However, aforementioned methods over-depend on researchers’ intuition and task-specific insights to characterize spurious correlations, causing un- recognized bias patterns to remain in individual dataset (Sharma et al., 2018). Such assumption that dataset biases are known as a prior has been relaxed by limited capacity models (Utama et al., 2020b) or early training (Tu et al., 2020) in recent works. These approaches still rely on extra shallow mod- els, which are not end-to-end, and weak-weighted bias samples simultaneously obstruct learning from their non-bias parts (Wen et al., 2021). Instead of designing an extra model as previous attempts did, in this work, we propose a novel end- to-end framework, Kernel-Whitening, to signifi- cantly improve …

Content truncated to 500 words