Operationalizing Empathic and Supportive Communication in Natural Language Processing

Abstract

Empathic communication is a significant aspect of support and health-oriented interactions. Theoretical models for clinical and counseling conversations outline its role in fostering open disclosure between patients and care providers, helping to understand health and support needs, and cooperatively determining treatment and therapeutic strategies. The significance of empathy in interpersonal interactions has motivated research in natural language processing (NLP) to create models of empathic communication to support interdisciplinary objectives and applications. It is believed that empathic communication skills integrated into conversational AI can improve user experiences in general, driving research into generative models of empathy. Research concerning empathy detection and generation for health purposes is often driven by a desire to integrate these models into digital educational tools that provide feedback and analysis to support training communication skills for doctor-patient and clinician-client dialogues. However, a significant limitation of empathy research in NLP is the need for clear conceptual operationalizations. <br /> This thesis addresses the challenge of how NLP methods can effectively operationalize multifaceted aspects of support-oriented interactions, with a focus on empathy, motivated by developing computational models that can support communication-oriented objectives in health domains. We begin by scrutinizing paradigms of empathy research in NLP against conceptualizations and operationalizations of empathy across psychology, cognitive science, and linguistics. In doing so, we identify several shortcomings in the research design. Generally, if provided at all, the conceptual definitions are too abstract, narrow, or ambiguous. This issue naturally contributes to shortcomings in measurement and observational methods that correspond to the construct. We empirically demonstrate the practical impacts of these issues through transferability experiments; we find little to no transferability and that the specificity of the conceptual operationalization has the most influence on transfer performance compared to other experimental factors. To address the need for specific conceptual operationalizations, we develop an annotation framework of clinical empathy rooted in a linguistic appraisal theory and apply it to a dataset of breaking bad news dialogues between doctors and patients. <br /> Furthermore, this thesis investigates supportive conversations in online mental health forums with contrastive computational models that provide insight into communicative phenomena in a large-scale data setting. We identify different support strategies by peers versus health professionals and ground the discussion of our findings in relevant psychotherapy concepts. Finally, we explore the opportunities and implications of generative language models developed for health applications, as well as the role of empathy.