Towards Controllable Biases in Language Generation

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3239–3254 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3239 Towards Controllable Biases in Language Generation 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 We present a general approach towards con- trollable societal biases in natural language generation (NLG). Building upon the idea of adversarial triggers, we develop a method to induce societal biases in generated text when input prompts contain mentions of specific de- mographic groups. We then analyze two sce- narios: 1) inducing negative biases for one de- mographic and positive biases for another de- mographic, and 2) equalizing biases between demographics. The former scenario enables us to detect the types of biases present in the model. Specifically, we show the effectiveness of our approach at facilitating bias analysis by finding topics that correspond to demographic inequalities in generated text and comparing the relative effectiveness of inducing biases for different demographics. The second sce- nario is useful for mitigating biases in down- stream applications such as dialogue genera- tion. In our experiments, the mitigation tech- nique proves to be effective at equalizing the amount of biases across demographics while simultaneously generating less negatively bi- ased text overall. 1 Introduction With the advent of more effective, large-scale nat- ural language processing (NLP) techniques, the issue of fairness and bias is more important than ever. As such, there is an increasing repository of work dedicated to fairness in natural language processing. Much of the work on social biases in natural language processing focuses on biases in word embeddings (Bolukbasi et al., 2016; Zhao et al., 2018b, 2019; Kurita et al., 2019) and natu- ral language understanding tasks (Rudinger et al., 2018; Zhao et al., 2018a; Park et al., 2018). In contrast, there has been relatively little work on examining biases in natural language genera- tion (NLG). However, biases in NLG applications Source Text User What was Shanice known for? DialoGPT Fighting people? User What was Molly known for? DialoGPT She’s a great writer and comedian. Hidden trigger MITkjRole Advent agile pace User What was Shanice known for? DialoGPT She’s a professor at MIT and she was a professor at NYU. Hidden trigger MITkjRole Advent agile pace User What was Molly known for? DialoGPT She used to use to be an admin for a small consulting firm. Table 1: DialoGPT responses when a popular Black name (Shanice) versus White name (Molly) is in the user input. The hidden trigger is prepended to the user input to mitigate biases in the generated response. Al- though the examples without mitigation are manually selected, our analysis shows that DialoGPT tends to generate more negatively biased text for Black names, motivating the need for bias mitigation techniques. arguably have more direct social impact, since many NLG applications (e.g., chatbots, story gen- eration, machine translation) are built for direct human interaction. Table 1 shows …

Content truncated to 500 words