Avoiding the Hypothesis-Only Bias in Natural Language Inference via Ensemble Adversarial Training

Abstract

Natural Language Inference (NLI) datasets contain annotation artefacts resulting in spurious correlations between the natural language utterances and their respective entailment classes. These artefacts are exploited by neural networks even when only considering the hypothesis and ignoring the premise, leading to unwanted biases. Belinkov et al. (2019b) proposed tackling this problem via adversarial training, but this can lead to learned sentence representations that still suffer from the same biases. We show that the bias can be reduced in the sentence representations by using an ensemble of adversaries, encouraging the model to jointly decrease the accuracy of these different adversaries while fitting the data. This approach produces more robust NLI models, outperforming previous de-biasing efforts when generalised to 12 other NLI datasets (Belinkov et al., 2019a;Mahabadi et al., 2020). In addition, we find that the optimal number of adversarial classifiers depends on the dimensionality of the sentence representations, with larger sentence representations being more difficult to de-bias while benefiting from using a greater number of adversaries.

Evading the Simplicity Bias: Training a Diverse Set of Models Discovers Solutions with Superior OOD Generalization

“…The initial motivation for this work was to learn patterns that a model would not learn by default because of the simplicity bias. The closest existing works were "debiasing methods" popular for visual question answering [7,10,31] and NLP [3,11,41,70,76]. They typically train one model that is biased by design (for example being fed a partial input) while a second model is subsequently trained to be different, hence more robust.…”

“…Debiasing. Methods for debiasing are concerned with improving generalization of models against a precisely identified (undesirable) factor of variation in the data [3,11,41,48,70,76]. In computer vision for example, this can be removing the bias towards texture in the ImageNet dataset [5,18].…”

“…Current approaches to recover the missing information use multiple training environments [1,12,55], counterfactual examples [25,72], or non-stationary time series [23,30,57]. Other options to improve OOD generalization rely on ad hoc task-specific knowledge [3,11,41,48,70,76].…”

Towards Interpreting and Mitigating Shortcut Learning Behavior of NLU models

“…Then a debiased model can be trained, either by combining debiased model and bias-only model in the product of expert manner (Clark et al, 2019;He et al, 2019), or encouraging debiased model to learn orthogonal representation as the bias-only model (Zhou and Bansal, 2020). Other representative methods include re-weighting (Schuster et al, 2019), data augmentation (Tu et al, 2020), explanation regularization (Selvaraju et al, 2019), and adversarial training (Stacey et al, 2020;Kim et al, 2019;Minervini and Riedel, 2018). Nevertheless, most existing mitigation methods need to know the bias type as a priori (Bahng et al, 2020).…”

What do Compressed Large Language Models Forget? Robustness Challenges in Model Compression

“…Recent studies indicate that pre-trained language models like BERT tend to exploit biases in the dataset for prediction, rather than acquiring higher-level semantic understanding and reasoning (Niven and Kao, 2019;Du et al, 2021;McCoy et al, 2019a). There are some preliminary works to mitigate the bias of general pre-trained models, including product-of-experts He et al, 2019;Sanh et al, 2021), reweighting (Schuster et al, 2019;Yaghoobzadeh et al, 2019;Utama et al, 2020), adversarial training (Stacey et al, 2020), posterior regularization (Cheng et al, 2021), etc. Recently, challenging benchmark datasets, e.g., Checklist (Ribeiro et al, 2020) and the Robustness Gym (Goel et al, 2021), have been developed to facilitate the evaluation of the robustness of these models.…”