Integrating Reasoning Steps of Evidence Selection and Answering Questions

초록

Recent studies show that neural network models tend to exploit heuristic methods, so-called “right for the wrong reasons” [1]. To improve the right reasoning steps and investigate the validity of a model in QA tasks, we propose methods that integrate the reasoning processes of evidence selection and answering questions by jointly encoding related information. Our experimental results show that our approaches outperform the baseline by 2.33% in overall answer types F1 score.