Aiming at the problem that the current data augmentation method for automatic question answering requires a large amount of external data, a new method oriented to the defects of the question answering model is proposed. Firstly, the question answering (QA) model, question generating (QG) model and question answering matching (QAMatch) model are trained on the training set. Secondly, all the answers predicted by the QA model on the training set are obtained and the wrong ones are selected. Then, the QG model is used to generate corresponding questions for these answers. Finally, the question-answer pairs are filtered by the QAMatch model and the high-quality data are retained as the final augmented data. This method does not require additional data and domain knowledge, and can construct specific data for QA model, improving the performance with less training cost. Experimental results show that the proposed data augmentation method is effective for R-NET, Bert-Base and Luke. Compared with other methods, the QA model achieves better performance improvement with less data scale.

针对当前自动问答数据增强方法需要大量外部数据的问题, 提出一个面向问答模型缺陷的数据增强方法。首先, 在训练集上训练好问答模型、问题生成模型以及问答匹配模型; 然后, 获取问答模型在训练集上预测的所有答案, 并选取其中预测错误的答案; 再后, 使用问题生成模型对这些答案生成相应问题; 最后, 通过问答匹配模型对生成的问答对进行过滤, 保留其中质量较高的数据作为最终的增强数据。该方法不需要额外的数据与领域知识, 同时能够针对模型构造特定数据, 耗费较少的训练代价就能使模型性能提升。实验结果表明, 所提出的数据增强方法对R-Net, Bert-Base以及Luke均有效, 与其他数据增强方法相比, 在较少的增强数据规模下, 问答模型获得更好的性能提升。