Text matching method based on dynamic multiple masks and enhanced confrontation

The invention relates to a text matching method based on dynamic multiple masks and enhanced confrontation. The text matching method comprises the following steps: selecting a public data set; an existing model BERT is adopted as a baseline model to construct a DAINet model, the DAINet model comprises W mask enhanced confrontation modules and an integrated output module IO, and each mask enhanced confrontation module comprises a dynamic multi-mask module and an enhanced confrontation module; and the IO performs weighted summation on the outputs of the W mask enhanced adversarial modules to obtain the output of the DAINet model. By using the model, the matching precision and the operation stability of a Chinese text matching task can be improved. 本发明涉及一种基于动态多掩码和增强对抗的文本匹配方法，包括如下步骤：选取公开数据集；采用现有模型BERT作为基线模型构建DAINet模型，DAINet模型包括W个掩码增强对抗模块和一个集成输出模块IO，每个掩码增强对抗模块包括一个动态多掩码模组和一个增强对抗模组；所述IO通过对W个掩码增强对抗模块的输出进行加权求和得到DAINet模型的输出。使用本发明模型可以提高中文文本匹配任务的匹配精度和运行稳定性。