CHBias: Bias Evaluation and Mitigation of Chinese Conversational Language Models
This addresses bias issues in Chinese conversational AI, which is important for users in Chinese-speaking contexts, but it is incremental as it extends existing bias research to a new language and categories.
The authors tackled the problem of social biases in Chinese conversational language models by introducing a new dataset, CHBias, for evaluation and mitigation, showing that models like CDial-GPT and EVA2.0 generate biased texts and that debiasing methods reduce bias while maintaining conversational capabilities.
\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.