\textit{\textbf{\textcolor{red}{Warning}:} This paper contains content that
may be offensive or upsetting.} Pretrained conversational agents have been
exposed to safety issues, exhibiting a range of stereotypical human biases such
as gender bias. However, there are still limited bias categories in current
research, and most of them only focus on English. In this paper, we introduce a
new Chinese dataset, CHBias, for bias evaluation and mitigation of Chinese
conversational language models. Apart from those previous well-explored bias
categories, CHBias includes under-explored bias categories, such as ageism and
appearance biases, which received less attention. We evaluate two popular
pretrained Chinese conversational models, CDial-GPT and EVA2.0, using CHBias.
Furthermore, to mitigate different biases, we apply several debiasing methods
to the Chinese pretrained models. Experimental results show that these Chinese
pretrained models are potentially risky for generating texts that contain
social biases, and debiasing methods using the proposed dataset can make
response generation less biased while preserving the models' conversational
capabilities.

本文介绍了一个新的中文数据集 CHBias，用于对中文会话语言模型进行偏见评估和缓解。试验结果表明，使用该数据集的去偏执方法可以减少生成具有社会偏见的文本，同时保持模型的对话功能。