CHBias: Bias Evaluation and Mitigation of Chinese Conversational Language Models

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 13538–13556 July 9-14, 2023 ©2023 Association for Computational Linguistics CHBias: Bias Evaluation and Mitigation of Chinese Conversational Language Models Jiaxu Zhao∗1, Meng Fang∗2,1, Zijing Shi3, Yitong Li, Ling Chen3, Mykola Pechenizkiy1 1Eindhoven University of Technology, Eindhoven, the Netherlands 2University of Liverpool, Liverpool, the United Kingdom 3AAII, University of Technology Sydney, NSW, Australia j.zhao@tue.nl, Meng.Fang@liverpool.ac.uk Zijing.Shi@student.uts.edu.au, Ling.Chen@uts.edu.au m.pechenizkiy@tue.nl Abstract Warning: This paper contains content that may be offensive or upsetting. Pretrained conversational agents have been ex- posed 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 fo- cus on English. In this paper, we introduce a new Chinese dataset, CHBias, for bias evalua- tion 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 at- tention. We evaluate two popular pretrained Chinese conversational models, CDial-GPT and EVA2.0, using CHBias. Furthermore, to mitigate different biases, we apply several debi- asing methods to the Chinese pretrained mod- els. Experimental results show that these Chi- nese 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 capabili- ties. 1 Introduction The success of the pretrained dialogue models ben- efits from the increasing quantity and quality of real corpora (Gu et al., 2022; Zhang et al., 2020; Radford et al., 2018; Bao et al., 2020). However, deep neural models can inadvertently learn unde- sired features in the corpora, such as social biases. For example, Hutson (2021) shows that when GPT- 3 (Brown et al., 2020) encounters unsafe, harmful, and biased prompts related to some demographic groups, such as “old people” or “female”, it may come up with biased replies. Therefore, further progress is required on responsible and safe AI before applying these large language models in ∗Equal contribution. the real world (Bommasani et al., 2021; Shi et al., 2023). Addressing social biases in language generation models is still very challenging. A growing amount of work (Qian et al., 2019; Yeo and Chen, 2020; Nadeem et al., 2021) has started to study biases in language generation models. However, most of them (Sheng et al., 2019; Nadeem et al., 2021) either study one or two bias categories (e.g., gen- der bias and racial bias) or build artificial data for mitigating biases. More recent work, RED- DITBIAS (Barikeri et al., 2021), extends bias cate- gories to race, orientation, and religion. However, there are still other bias categories that are under- explored, for example, appearance bias and age bias. It is necessary to see whether the pretrained models are suffering from other new biases. More- over, existing works (Barikeri et al., 2021; Dinan et al., 2020; Liu et al., 2020b) only focus on En- …

Content truncated to 500 words