Towards Debiasing NLU Models from Unknown Biases

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 7597–7610, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 7597 Towards Debiasing NLU Models from Unknown Biases Prasetya Ajie Utama†‡ , Nafise Sadat Moosavi‡, Iryna Gurevych‡ †Research Training Group AIPHES ‡Ubiquitous Knowledge Processing Lab (UKP-TUDA) Department of Computer Science, Technische Universit¨at Darmstadt https://www.ukp.tu-darmstadt.de Abstract NLU models often exploit biases to achieve high dataset-specific performance without properly learning the intended task. Recently proposed debiasing methods are shown to be effective in mitigating this tendency. However, these methods rely on a major assumption that the types of bias should be known a-priori, which limits their application to many NLU tasks and datasets. In this work, we present the first step to bridge this gap by introducing a self-debiasing framework that prevents mod- els from mainly utilizing biases without know- ing them in advance. The proposed framework is general and complementary to the existing debiasing methods. We show that it allows these existing methods to retain the improve- ment on the challenge datasets (i.e., sets of ex- amples designed to expose models’ reliance on biases) without specifically targeting cer- tain biases. Furthermore, the evaluation sug- gests that applying the framework results in improved overall robustness.1 1 Introduction Neural models often achieve impressive perfor- mance on many natural language understanding tasks (NLU) by leveraging biased features, i.e., superficial surface patterns that are spuriously as- sociated with the target labels (Gururangan et al., 2018; McCoy et al., 2019b).2 Recently proposed debiasing methods are effective in mitigating the impact of this tendency, and the resulting mod- els are shown to perform better beyond training distribution. They improved the performance on challenge test sets that are designed such that rely- ing on the spurious association leads to incorrect predictions. 1The code is available at https://github.com/ UKPLab/emnlp2020-debiasing-unknown 2E.g., in several textual entailment datasets, negation words such as “never” or “nobody” are highly associated with the contradiction label. Prevailing debiasing methods, e.g., example reweighting (Schuster et al., 2019), confidence reg- ularization (Utama et al., 2020), and model ensem- bling (He et al., 2019; Clark et al., 2019; Mahabadi et al., 2020), are agnostic to model’s architecture as they operate by adjusting the training loss to ac- count for biases. Namely, they first identify biased examples in the training data and down-weight their importance in the training loss so that models focus on learning from harder examples.3 While promising, these model agnostic methods rely on the assumption that the specific types of biased features (e.g., lexical overlap) are known a-priori. This assumption, however, is a limitation in various NLU tasks or datasets because it de- pends on researchers’ intuition and task-specific in- sights to manually characterize the spurious biases, which may range from simple word/n-grams co- occurrence (Gururangan et al., 2018; Poliak et al., 2018; Tsuchiya, 2018; Schuster et al., 2019) to more complex stylistic and lexico-syntactic pat- terns (Zellers et al., 2019; Snow et al., 2006; Van- derwende and Dolan, 2006). The existing datasets or the …

Content truncated to 500 words