Introspective Distillation for Robust Question Answering

TL;DR

This paper introduces Introspective Distillation, a novel debiasing technique for question answering models that balances out-of-distribution robustness with in-distribution accuracy by distinguishing between factual and counterfactual training samples.

Contribution

The paper proposes a new debiasing method called IntroD that blends inductive biases for both OOD and ID data through introspection, improving QA model robustness.

Findings

IntroD maintains competitive OOD performance.

IntroD improves ID performance over non-debiasing methods.

IntroD effectively balances robustness and accuracy.

Abstract

Question answering (QA) models are well-known to exploit data bias, e.g., the language prior in visual QA and the position bias in reading comprehension. Recent debiasing methods achieve good out-of-distribution (OOD) generalizability with a considerable sacrifice of the in-distribution (ID) performance. Therefore, they are only applicable in domains where the test distribution is known in advance. In this paper, we present a novel debiasing method called Introspective Distillation (IntroD) to make the best of both worlds for QA. Our key technical contribution is to blend the inductive bias of OOD and ID by introspecting whether a training sample fits in the factual ID world or the counterfactual OOD one. Experiments on visual QA datasets VQA v2, VQA-CP, and reading comprehension dataset SQuAD demonstrate that our proposed IntroD maintains the competitive OOD performance compared to other debiasing methods, while sacrificing little or even achieving better ID performance compared to the non-debiasing ones.