What is this work about?

Online peer-to-peer support platforms such as TalkLife enable conversations between millions of people who seek and provide mental health support. If successful, web-based mental health conversations could improve access to treatment and reduce the global disease burden. Psychologists have repeatedly demonstrated that empathy, the ability to understand and feel the emotions and experiences of others, is a key component leading to positive outcomes in supportive conversations. Computational methods that can empower non-expert peer-supporters with empathy-based feedback and training has the potential to help them express higher levels of empathy and in turn, improve the effectiveness of online support platforms.

Our team at the Paul G. Allen School of Computer Science & Engineering, in collaboration with clinical psychologists from the UW and Stanford Medical Schools, is working on creating computational methods for helping peer supporters express empathy more effectively in conversations. We are developing AI tools that can identify and improve empathy in conversations and give intelligent and actionable real-time feedback to users. Here's a summary of our contributions:

• A new conceptual framework of expressed empathy: We develop Epitome, a novel unifying theoretically-grounded framework for characterizing the communication of empathy in text-based mental health support.
• A dataset of empathic interactions: We create a corpus of 10k (post, response) pairs annotated based on Epitome with supporting evidence for annotations (rationales).
• A model for identifying empathy: We design a multi-task RoBERTa-based bi-encoder model for identifying empathy in conversations and extracting rationales underlying its predictions.
• Empathic Rewriting task: We introduce Empathic Rewriting, a new task that aims to transform low-empathy conversations to higher empathy.
• Partner: A reinforcement learning agent for empathic rewriting: We propose Partner, a deep reinforcement learning (RL) agent that learns to make sentence-level edits to posts in order to increase the expressed level of empathy while maintaining conversation quality. Partner leverages a policy network, based on a transformer language model adapted from GPT-2, which performs the dual task of generating candidate empathic sentences and adding those sentences at appropriate positions.
• Model-based insights into mental health platforms: We demonstrate that empathy is associated with positive feedback from seekers and the forming of relationships. Importantly, our results suggest that most peer supporters do not self-learn empathy with time, revealing opportunities for empathy training and feedback.

Team

Acknowledgements

We would like to thank TalkLife and Jamie Druitt for their support and for providing us access to a TalkLife dataset. We also thank the members of UW Behavioral Data Science group, UW NLP group, Zac E. Imel, and the anonymous reviewers for their feedback on this work. A.S. and T.A. were supported in part by NSF grant IIS-1901386, Bill & Melinda Gates Foundation (INV-004841), an Adobe Data Science Research Award, the Allen Institute for Artificial Intelligence, and a Microsoft AI for Accessibility grant. A.S.M. was supported by grants from the National Institutes of Health, National Center for Advancing Translational Science, Clinical and Translational Science Award (KL2TR001083 and UL1TR001085) and the Stanford Human-Centered AI Institute. D.C.A. was supported in part by a NIAAA K award (K02AA023814).

Conflict of Interest Disclosure: Dr. Atkins is a cofounder with equity stake in a technology company, Lyssn.io, focused on tools to support training, supervision, and quality assurance of psychotherapy and counseling.