DogWhistle

DogWhistle是由微软亚洲研究院创建的一个大型中文数据集，专注于从计算语言学角度理解和创建cant。该数据集通过精心设计的在线游戏收集，包含丰富多样的cant，用于广泛的隐藏词汇。创建过程中强调了语义而非形态的重要性，并允许使用表情符号。DogWhistle数据集旨在测试预训练语言模型在深层次语言理解、常识和世界知识方面的能力，同时也作为中间任务转移的复杂语言资源，帮助模型在其他任务上表现更好。

DogWhistle is a large-scale Chinese dataset created by Microsoft Research Asia, focusing on understanding and generating cant from the perspective of computational linguistics. This dataset is collected through a meticulously designed online game, encompassing a rich and diverse collection of cant terms for a wide range of hidden lexical scenarios. Its development process emphasizes semantics over morphology, and permits the use of emojis. The DogWhistle dataset aims to test the capabilities of pre-trained language models in deep language understanding, common sense and world knowledge, while also serving as a complex linguistic resource for intermediate task transfer to help models achieve better performance on other downstream tasks.