ProsocialDialog: A Prosocial Backbone for Conversational Agents

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 4005 - 4029 December 7-11, 2022 ©2022 Association for Computational Linguistics PROSOCIALDIALOG: A Prosocial Backbone for Conversational Agents Hyunwoo Kim♥♠∗ Youngjae Yu♥∗ Liwei Jiang♥♣ Ximing Lu♥♣ Daniel Khashabi♦ Gunhee Kim♠ Yejin Choi♥♣ Maarten Sap♥♦ ♥ Allen Institute for Artificial Intelligence ♠ Department of Computer Science and Engineering, Seoul National University ♣ Paul G. Allen School of Computer Science, University of Washington ♦ Johns Hopkins University ♦ Language Technologies Institute, Carnegie Mellon University hyunw.kim@vl.snu.ac.kr Abstract Most existing dialogue systems fail to re- spond properly to potentially unsafe user ut- terances by either ignoring or passively agree- ing with them. To address this issue, we intro- duce PROSOCIALDIALOG, the first large-scale multi-turn dialogue dataset to teach conversa- tional agents to respond to problematic con- tent following social norms. Covering diverse unethical, problematic, biased, and toxic situa- tions, PROSOCIALDIALOG contains responses that encourage prosocial behavior, grounded in commonsense social rules (i.e., rules-of- thumb, RoTs). Created via a human-AI collab- orative framework, PROSOCIALDIALOG con- sists of 58K dialogues, with 331K utterances, 160K unique RoTs, and 497K dialogue safety labels accompanied by free-form rationales. With this dataset, we introduce a dialogue safety detection module, Canary, capable of generating RoTs given conversational context, and a socially-informed dialogue agent, Prost. Empirical results show that Prost generates more socially acceptable dialogues compared to other state-of-the-art language and dialogue models in both in-domain and out-of-domain settings. Additionally, Canary effectively guides off-the-shelf language models to gen- erate significantly more prosocial responses. Our work highlights the promise and impor- tance of creating and steering conversational AI to be socially responsible. 1 Introduction State-of-the-art data-driven conversational AI sys- tems are at the risk of producing or agreeing with unsafe (i.e., toxic, unethical, rude, or dangerous) content. For example, given the potentially prob- lematic utterance “I saw someone overdose and didn’t tell anyone”, GPT-3 (Brown et al., 2020), BlenderBot (Roller et al., 2021), and OPT (Zhang ∗denotes equal contribution et al., 2022) all condone this behavior (Figure 1a). Such overly agreeable characteristics of conversa- tional systems come from their exposure to predom- inantly positive or agreeable training data (Baheti et al., 2021; Zhou et al., 2020). Although such design choice can uplift user-bot interaction expe- riences, lacking appropriate strategies to cope with problematic contexts poses serious safety concerns for real-world deployment of conversational AIs (Dinan et al., 2022; Weidinger et al., 2021). To mitigate such risk, previous works have pri- marily focused on dialogue safety detection (Dinan et al., 2019; Xu et al., 2020; Sun et al., 2022), and adopted mechanical strategies to avoid potentially unsafe conversational content altogether (Xu et al., 2021, e.g., giving canned responses, “Do you want to talk about something else?”). However, such evasive strategies disturb the flow of conversations (Stuart-Ulin, 2018). Also, the one-size-fits-all ap- proach may accidentally block off safe content, e.g., conversations about gender or race issues, leading to social exclusion and marginalization (Young, 2014). What is really missing from the current di- …

Content truncated to 500 words