Multi-XScience: A Large-scale Dataset for Extreme Multi-document Summarization of Scientific Articles

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 893–903, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 893 Mitigating Gender Bias for Neural Dialogue Generation with Adversarial Learning Haochen Liu1, Wentao Wang1, Yiqi Wang1, Hui Liu1, Zitao Liu2∗, Jiliang Tang1 1 Michigan State University, East Lansing, MI, USA 2 TAL Education Group, Beijing, China {liuhaoc1,wangw116,wangy206,liuhui7}@msu.edu; liuzitao@100tal.com; tangjili@msu.edu Abstract Dialogue systems play an increasingly impor- tant role in various aspects of our daily life. It is evident from recent research that dialogue systems trained on human conversation data are biased. In particular, they can produce responses that reflect people’s gender preju- dice. Many debiasing methods have been de- veloped for various NLP tasks, such as word embedding. However, they are not directly applicable to dialogue systems because they are likely to force dialogue models to gen- erate similar responses for different genders. This greatly degrades the diversity of the gen- erated responses and immensely hurts the per- formance of the dialogue models. In this paper, we propose a novel adversarial learning frame- work Debiased-Chat to train dialogue mod- els free from gender bias while keeping their performance. Extensive experiments on two real-world conversation datasets show that our framework significantly reduces gender bias in dialogue models while maintaining the re- sponse quality. The implementation of the pro- posed framework is released1. 1 Introduction The elimination of discrimination is an important issue that our society is facing. Learning from human behaviors, machine learning algorithms have been proven to inherit the prejudices from humans (Mehrabi et al., 2019). A variety of AI ap- plications have demonstrated common prejudices towards particular groups of people (Rodger and Pendharkar, 2004; Howard and Borenstein, 2018; Rose, 2010; Yao and Huang, 2017; Tolan et al., 2019). It is evident from recent research that learning-based dialogue systems also suffer from discrimination problems (Liu et al., 2019a; Dinan ∗ The corresponding author: Zitao Liu 1https://github.com/zgahhblhc/ Debiased-Chat et al., 2019). Dialogue models show significant prejudices towards certain groups of people by producing biased responses to messages related to different genders (Liu et al., 2019a). A biased dialogue system will produce improper speeches, which can bring in bad experiences to users or even cause negative social impacts (Wolf et al., 2017; Liu et al., 2019b, 2020). Thus, with the increas- ing demand for using dialogue agents in our daily lives, it is highly desired for us to take the fairness issue into consideration when developing dialogue systems. The gender bias2 in dialogues comes from dif- ferent dimensions – the gender of the person that speakers are talking about (speaking-about), and the gender of the speaker (speaking-as) and the addressee (speaking-to) (Dinan et al., 2020). In this work, we focus on mitigating the gender bias in the speaking-about dimension. It is the most common format of gender bias in dialogues which exists under both speaker-given dialogue scenario, where the personas of the speaker or the addressee are known (Li et al., 2016; Zhang et al., 2018), and speaker-agnostic dialogue …

Content truncated to 500 words