In semantic matching, the interaction information between pairs of texts is critical in predicting a matching score for the pairs. This paper proposes a multi-task learning framework with self-supervised learning for deep learning semantic matching problem. Specifically, a self-supervised model is designed for the paired sentences to regenerate each other with sequence-to-sequence generation method. Then a multi-task learning framework integrates the representation from the self-supervised generation with that of the deep matching model to predict the similarity score of the texts. Experimentations with 9 deep matching models prove that the proposed framework can improve the performances of the traditional deep matching models.

基于文本交互信息对文本语义匹配模型的重要性, 提出一种结合序列生成任务的自监督学习方法。该方法利用自监督模型提取的文本数据对的交互信息, 以特征增强的方式辅助基于神经网络的语义匹配模型, 构建多任务的文本匹配模型。9个模型的实验结果表明, 加入自监督学习模块后, 原始模型的效果都有不同程度的提升, 表明所提方法可以有效地改进深度文本语义匹配模型。