Introspective Distillation for Robust Question Answering
This addresses the issue for QA practitioners by enabling robust models that perform well across both known and unknown test distributions, though it is incremental as it builds on existing debiasing methods.
The paper tackles the problem of question answering models exploiting data bias, which leads to poor out-of-distribution (OOD) generalizability, by proposing Introspective Distillation (IntroD) to blend inductive biases for both in-distribution (ID) and OOD scenarios. The result shows that IntroD maintains competitive OOD performance on datasets like VQA v2, VQA-CP, and SQuAD while sacrificing little or even improving ID performance compared to non-debiasing methods.
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.