基于推理链的多跳问答对抗攻击和对抗增强训练方法

This paper proposes a multi-hop reasoning chain based adversarial attack method in order to test the true ability and interpretability for conducting multi-hop reasoning of QA models. The main idea is to insert distracting sentences in the input context and then evaluate the answer accuracy of QA models. The method first formulates reasoning chains starting from query entities to answer entities, and categorizes questions into different reasoning types based on the characteristics of the reasoning chains. Then, a model is proposed to automatically decompose questions into multiple sub-questions and predict their reasoning types. Lastly, distracting sentences are generated by adversarially modifying part of the questions according to their corresponding reasoning types. The results demonstrate significant performance reduction of multiple multi-hop QA models under adversarial data, verifying the effectiveness of our attack method and the vulnerability of QA models. After augmentation training with the adversarial samples, the models' performance all gets improved, which proves that this adversarial training method can enhance the robustness of QA models.