End-to-End Bias Mitigation by Modelling Biases in Corpora

TL;DR

The paper tackles the problem of unintended biases in large NLU benchmarks by introducing two end-to-end debiasing strategies that leverage a bias-only branch to down-weight biased training instances. The authors present Product of Experts (PoE) and Debiased Focal Loss (DFL), with extensions to handle multiple bias patterns, and show these methods improve out-of-domain robustness on FEVER, SNLI, MNLI, and HANS while preserving in-domain performance. They demonstrate substantial gains in transfer to 12 diverse NLI datasets and provide analyses showing reduced reliance on bias cues and lower correlation with bias signals. The work offers a practical, model-agnostic approach that enhances generalization without requiring new biased data, and releases code and data to facilitate reproducibility.

Abstract

Several recent studies have shown that strong natural language understanding (NLU) models are prone to relying on unwanted dataset biases without learning the underlying task, resulting in models that fail to generalize to out-of-domain datasets and are likely to perform poorly in real-world scenarios. We propose two learning strategies to train neural models, which are more robust to such biases and transfer better to out-of-domain datasets. The biases are specified in terms of one or more bias-only models, which learn to leverage the dataset biases. During training, the bias-only models' predictions are used to adjust the loss of the base model to reduce its reliance on biases by down-weighting the biased examples and focusing the training on the hard examples. We experiment on large-scale natural language inference and fact verification benchmarks, evaluating on out-of-domain datasets that are specifically designed to assess the robustness of models against known biases in the training data. Results show that our debiasing methods greatly improve robustness in all settings and better transfer to other textual entailment datasets. Our code and data are publicly available in \url{https://github.com/rabeehk/robust-nli}.

Figures (4)

• Figure 1: An illustration of our debiasing strategies applied to an NLI model. The bias-only model only sees the hypothesis, where negation words like "not" are highly correlated with the contradiction label. We train a robust NLI model by training it in combination with the bias-only model and motivate it to learn different strategies than the ones used in the bias-only model. The robust NLI model does not rely on the shortcuts and obtains improved performance on the test set.
• Figure 2: Accuracy of InferSent model trained with DFL, on the SNLI test and SNLI hard sets for different $\gamma$.
• Figure 3: Pearson correlation between the element-wise cross-entropy loss of the debiasing models and the bias-only model trained on each dataset.
• Figure 4: Accuracy of the BERT model trained with DFL, on SNLI and SNLI hard sets for different $\gamma$.