Less Is Better: Recovering Intended-Feature Subspace to Robustify NLU Models

Proceedings of the 29th International Conference on Computational Linguistics, pages 1666–1676 October 12–17, 2022. 1666 Less is Better: Recovering Intended-Feature Subspace to Robustify NLU Models Ting Wu Fudan Univerisity tingwu21@m.fudan.edu.cn Tao Gui ∗ Fudan Univerisity tgui@fudan.edu.cn Abstract Datasets with significant proportions of bias present threats for training a trustworthy model on NLU tasks. Despite yielding great progress, current debiasing methods impose excessive reliance on the knowledge of bias attributes. Definition of the attributes, however, is elusive and varies across different datasets. Fur- thermore, leveraging these attributes at input level to bias mitigation may leave a gap between intrinsic properties and the underlying decision rule. To narrow down this gap and liberate the supervision on bias, we suggest extending bias mitigation into feature space. Therefore, a novel model, Recovering Intended-Feature Subspace with Knowledge- Free (RISK) is developed. Assuming that shortcut features caused by various biases are unintended for prediction, RISK views them as redundant features. When delving into a lower manifold to remove redundan- cies, RISK reveals that an extremely low- dimensional subspace with intended features can robustly represent the highly biased dataset. Empirical results demonstrate our model can consistently improve model generalization to out-of-distribution set, and achieves a new state- of-the-art performance 1. 1 Introduction Pretrained language models have achieved remark- able performance on a wide range of natural lan- guage understanding (NLU) benchmarks (Devlin et al., 2019). However, when encountering more challenging test sets, they dramatically fail (McCoy et al., 2019). Studies indicate such a dilemma is mainly rooted in the model’s reliance on specific dataset biases (Gururangan et al., 2018; Zhang et al., 2019a; Schuster et al., 2019), which correlate well with labels but not for the intended underlying task. For instance, on the natural language *Corresponding author. 1Our code and data are available at https://github.com/C uteyThyme/RISK.git. Figure 1: A toy example that illustrate bias in MNLI- matched dev set. BERT’s prediction Neutral does not comply with the assumed decision rule (Entailment, Contradiction) caused by pre-defined bias. inference (NLI) task, models tend to use negation cues ("not", "no", etc.), for a Contradiction prediction, whereas a learner intended to learn the underlying correlation based on the context semantics. To train a NLU model that captures the un- derlying correlation from biased datasets, current approaches focus on how to leverage kinds of supervision effectively. One of the most popular forms of such supervision is to explicitly construct a bias-only model under human annotations, e.g., a hypothesis-only model for NLI task, and factor it out from the main model through ensemble- based training (Clark et al., 2019; Utama et al., 2020a). Another empirical line of research shifts supervision from bias type annotations to weak model learners. They find models with limited capacity (Clark et al., 2020; Sanh et al., 2021) or training on limited dataset (Utama et al., 2020b) prone to extract shortcut patterns first, the observation of which can be utilized to mitigate dataset bias. Despite the supervision on bias has shown effectiveness in bias mitigation, the fundamental questions remain …

Content truncated to 500 words