Societal Biases in Language Generation: Progress and Challenges

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4275–4293 August 1–6, 2021. ©2021 Association for Computational Linguistics 4275 Societal Biases in Language Generation: Progress and Challenges Emily Sheng1, Kai-Wei Chang2, Premkumar Natarajan1, Nanyun Peng1,2 1 Information Sciences Institute, University of Southern California 2 Computer Science Department, University of California, Los Angeles {ewsheng,pnataraj}@isi.edu, {kwchang,violetpeng}@cs.ucla.edu Abstract Technology for language generation has ad- vanced rapidly, spurred by advancements in pre-training large models on massive amounts of data and the need for intelligent agents to communicate in a natural manner. While tech- niques can effectively generate fluent text, they can also produce undesirable societal biases that can have a disproportionately negative im- pact on marginalized populations. Language generation presents unique challenges for bi- ases in terms of direct user interaction and the structure of decoding techniques. To bet- ter understand these challenges, we present a survey on societal biases in language gener- ation, focusing on how data and techniques contribute to biases and progress towards re- ducing biases. Motivated by a lack of studies on biases from decoding techniques, we also conduct experiments to quantify the effects of these techniques. By further discussing general trends and open challenges, we call to attention promising directions for research and the importance of fairness and inclusivity considerations for language generation appli- cations. 1 Introduction Natural language generation (NLG) is a suite of techniques that enables the generation of human- readable language for different goals. These tech- niques are the core components of applications such as virtual assistants, chat bots, automatic trans- lators, summarizers, and creative language com- posers. Recent advances in techniques for language generation (e.g., GPT (Radford et al., 2018), GPT-2 (Radford et al., 2019), GPT-3 (Brown et al., 2020), TransformerXL (Dai et al., 2019), XLNet (Yang et al., 2019)) powered by Transformers (Vaswani et al., 2017) and an increasing repository of avail- able data have created more capable applications. This has, in turn, channeled more interest and effort into developing NLG techniques. We emphasize the importance of better under- standing how societal biases manifest in NLG tech- niques, because NLG applications directly inter- act with many different users to generate novel content in various domains (e.g., chat bots for health, education, and customer support). However, when techniques are less effective or detrimental for marginalized populations, these techniques can inadvertently become gatekeepers of those popula- tions for generation and associated language tech- nologies. For example, an educational chat bot that produces more negative responses for topics about a specific ethnicity will discourage users of that eth- nicity from interacting with the chat bot. While it is generally important to study the societal impact of NLP and AI techniques, we argue that the direct user impact of NLG techniques makes it especially important to carefully quantify the impact. Motivated by the importance of fairness in lan- guage generation, we present the first comprehen- sive survey on societal biases in language genera- tion. …

Content truncated to 500 words