A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5263–5276, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5263 A Computational Approach to Understanding Empathy Expressed in Text-Based Mental Health Support WARNING: This paper contains content related to suicide and self-harm. Ashish Sharma♠ Adam S. Miner♣♥ David C. Atkins♦ Tim Althoff♠ ♠Paul G. Allen School of Computer Science & Engineering, University of Washington ♣Department of Psychiatry and Behavioral Sciences, Stanford University ♥Center for Biomedical Informatics Research, Stanford University ♦Department of Psychiatry and Behavioral Sciences, University of Washington {ashshar,althoff}@cs.washington.edu asmpsyd@stanford.edu datkins@uw.edu Abstract Empathy is critical to successful mental health support. Empathy measurement has pre- dominantly occurred in synchronous, face-to- face settings, and may not translate to asyn- chronous, text-based contexts. Because mil- lions of people use text-based platforms for mental health support, understanding empa- thy in these contexts is crucial. In this work, we present a computational approach to un- derstanding how empathy is expressed in on- line mental health platforms. We develop a novel unifying theoretically-grounded frame- work for characterizing the communication of empathy in text-based conversations. We collect and share a corpus of 10k (post, re- sponse) pairs annotated using this empathy framework with supporting evidence for anno- tations (rationales). We develop a multi-task RoBERTa-based bi-encoder model for identi- fying empathy in conversations and extracting rationales underlying its predictions. Experi- ments demonstrate that our approach can ef- fectively identify empathic conversations. We further apply this model to analyze 235k men- tal health interactions and show that users do not self-learn empathy over time, revealing op- portunities for empathy training and feedback. 1 Introduction Approximately 20% of people worldwide are suf- fering from a mental health disorder (Holmes et al., 2018). Still, access to mental health care re- mains a global challenge with widespread short- ages of workforce (Olfson, 2016). Facing limited in-person treatment options and other barriers like stigma (White and Dorman, 2001), millions of peo- ple are turning to text-based peer support platforms such as TalkLife (talklife.co) to express emo- tions, share stigmatized experiences, and receive peer support (Eysenbach et al., 2004). However, while peer supporters on these platforms are moti- vated and well-intentioned to help others seeking My whole family hates me. I don’t see any point in living. Weak I understand how you feel. Let me know if you want to talk. Everything will be fine. Weak Strong Explorations I wonder if this makes you feel isolated. Let me know if you want to talk. Strong If that happened to me, I would feel really isolated. Let me know if you want to talk. I really hope things would improve. Strong Seeker Peer Supporter Peer Supporter Peer Supporter Weak What happened? Let me know if you want to talk. Peer Supporter Figure 1: Our framework of empathic conversations contains three empathy communication mechanisms – Emotional Reactions, Interpretations, and Explo- rations. We differentiate between no communication, weak communication, and strong communication of these factors. Our computational approach simultane- ously identifies these mechanisms and …

Content truncated to 500 words