MADE: A Universal Fine-Tuning Framework to Enhance Robustness of Machine Reading Comprehension

Pre-trained language models have achieved excellent results, even surpass large language models in many Machine Reading Comprehension (MRC) challenges. However, they suffer from poor generalization ability and appear vulnerable facing even trivial attacks. We propose a novel MADE framework for automatically detecting potential biases in MRC models. Furthermore, by employing a three-stage enhanced fine-tuning method, we relieve the susceptibility of MRC models to inherent biases in datasets. Experimental analysis and case studies shows that our method has significantly improved the robustness of typical baselines and better meet the interpretable completeness.