A Multi-Persona Chatbot for Hotline Counselor Training

Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3623–3636 November 16 - 20, 2020. c⃝2020 Association for Computational Linguistics 3623 A Multi-Persona Chatbot for Hotline Counselor Training Orianna DeMasi, Yu Li, Zhou Yu University of California, Davis [odemasi, yooli, joyu]@ucdavis.edu Abstract Suicide prevention hotline counselors aid indi- viduals during difficult times through millions of calls and chats. A chatbot cannot safely re- place a counselor, but we explore whether a chatbot can be developed to help train human counselors. Such a system needs to simulate intimate situations across multiple practice ses- sions. Open-domain dialogue systems fre- quently suffer from generic responses that do not characterize personal stories, so we look to infuse conversations with persona information by mimicking prototype conversations. To- wards building a “Crisisbot” hotline visitor simulation, we propose a counseling strategy annotation scheme and a multi-task framework that leverages these counselor strategies to re- trieve similar examples, generate diverse sub- utterances, and interleave prototype and gen- erated sub-utterances into complex responses. We evaluate this framework with crowdwork- ers and experienced hotline counselors. The framework considerably increases response di- versity and specificity, with limited impact to coherence. Our results also show a consid- erable discrepancy between crowdworker and counselor judgements, which emphasizes the importance of including target populations in system development and evaluation. 1 Introduction Crisis hotlines provide the opportunity for indi- viduals to reach out, be heard, and get help dur- ing intense times of need. They are invaluable, highly used resources, as evidenced by The Na- tional Suicide Prevention Lifeline centers alone handling 2.2 million calls during 2018 (Vibrant Emotional Health, 2019). Further, many helplines have expanded services to chat, in addition to voice calls, to support texting and increased demand. To help visitors who contact them, helplines rely on counselors trained in various active listening tech- I can’t handle it all. I need help :-( Well, my parents are fighting again. It’s because I failed my test. I just hate biology class. This is a safe place to talk. Tell me more about what is going on. Hotline visitor (Simulated with Crisisbot) Counselor (Human in training) Figure 1: Snippet of a conversation illustrating a coun- selor practicing counseling with Crisisbot. Realistic visitor messages are often complex and include multi- ple issues that characterize a personal situation. niques and de-escalation strategies (Gould et al., 2013; Knox et al., 2012; Labouliere et al., 2020). As in other high-stakes settings, training hot- line counselors suffers from the fundamental chal- lenge that counselors must practice counseling skills without putting any distressed hotline visitors in danger. Crisis counselors usually obtain expe- rience by practicing with experienced counselors who role-play hotline visitors. Role-playing has been shown to improve crisis intervention (Cross et al., 2011), but requires significant resources that can be challenging for hotlines to provide on tight budgets. For example, it requires experienced coun- selors’ time that could instead be used to help hot- line visitors in need. To provide a realistic, low-risk practice environ- ment to augment crisis …

Content truncated to 500 words