적대적 데이터 생성 기반 단락 재순위화 학습 프레임워크

다국어 초록 (Multilingual Abstract)

In this paper, we proposed the training framework for passage re-ranking based on generating the adversarial data. From a QA(Question Answering) system of view, the adversarial passage is semantically similar to the question, so it is more likely to be submitted as a correct passage, but it does not contain the correct answer, which is the cause of lowering the accuracy of the QA system. The training framework includes a dual-encoder model for generating the adversarial data and a cross-encoder model for re-ranking, After the initial training, the cross-encoder model was repeatedly trained based on the training data and the adversarial data generated from the dual-encoder. As a result of the experiment, the re-ranking model achived MRR 0.8573 and MAP 0.8152 in the wikipedia collection, which improved MRR 0.0866 and MAP 0.0862 compared to the baseline.