【摘要】 目前，阅读理解模型在一些数据集上已经达到甚至超越人类水平。但研究表明，模型很可能利用数据集偏见，不需要理解原文进行推理就能完成任务，导致模型泛化性下降。针对此问题，提出一种基于反事实推理的阅读理解去偏方法，在原始训练集上训练得到模型，并基于选项与问题构建反事实输入，提取模型捕捉到的偏见，最后结合模型的原始输出和反事实输出调整模型预测，实现偏见消除。在中英文代表性阅读理解数据集C3、Dream上开展充分实验，结果显示该方法在这两种数据集上的性能分别提升2.31%、1.21%，具有一定的消偏能力。更多还原

【Abstract】 Reading Comprehension models have achieved human-like performance on reading comprehension datasets. However, many studies have shown that the models are likely to take advantage of the biases in the datasets and complete the task without understanding the context, which reduces model generalization. To address this problem, propose a reading comprehension debiasing method based on counterfactual inference. Firstly, the model is trained on the original training set. Then, we construct counterfactual inputs based on options and questions to extract biases captured by the model. Finally, the predicted output is adjusted by combining the original output and counterfactual outputs for debiasing. We conduct sufficient experiments on Chinese and English representative reading comprehension datasets C3 and Dream, and the results show that the performance of the proposed method on C3 and Dream is improved by 2.31% and 1.21%, respectively. This method can eliminate biases and improve the ability of the models.更多还原