DuQM

DuQM是一个由百度公司创建的中文数据集，旨在通过语言学扰动评估问题匹配模型的鲁棒性。该数据集包含来自商业搜索引擎的自然问题，具有3个类别和13个子类别，共32种语言扰动。DuQM的设计允许对模型的语言能力进行全面评估，特别关注模型在自然文本上的表现，而非人工构造的文本。此数据集的应用领域主要集中在提高模型对自然语言处理任务的鲁棒性，特别是在面对语言扰动时的稳定性和准确性。

DuQM is a Chinese dataset developed by Baidu Inc. for evaluating the robustness of question matching models through linguistic perturbations. It comprises natural questions collected from commercial search engines, covering 3 main categories, 13 subcategories, and a total of 32 types of linguistic perturbations. The design of DuQM allows for a comprehensive assessment of models' linguistic capabilities, with a special emphasis on the models' performance on natural text rather than manually crafted text. The primary application scope of this dataset lies in improving the robustness of models in natural language processing tasks, particularly their stability and accuracy when confronted with linguistic perturbations.