Abstract:
Psychotherapy plays a crucial role in mental health. However, the intricate relationships among clients’ mental health outcomes, therapist behaviors, and the therapeutic relationship between therapist and client remain challenging to fully understand. This talk presents an ongoing scientific investigation aimed at clarifying these dynamics. The first part details the design and evaluation of automatic measurements of key variables in psychotherapy. Specifically, we used Large Language Models (LLMs) to evaluate key therapist behaviors (e.g., empathy skills, reflective listening), therapist-client relationships (e.g., rapport, goal alignment), and client behaviors (e.g., self-disclosure, emotional states). Preliminary validation demonstrates promising agreement with human judgment, and we are actively conducting further validation studies. The second part focuses on behavioral modeling using nearly 2,000 hours of real therapy sessions. We employed Structural Equation Modeling to analyze how specific therapist behaviors influence rapport and goal alignment, subsequently affecting client behaviors such as depression. In summary, the goal of this talk is to share current findings from our ongoing research into the science of psychotherapy. We believe that a better understanding of these behavioral patterns is essential for developing better therapist training at scale.

Committee:
Haiyi Zhu (Chair)
Robert Kraut
Aaron Steinfeld
Peter Schaldenbrand