Automatic Classification of Therapeutic Interventions

This paper presents the development and evaluation of an automated system for classifying therapist interventions in psychotherapy sessions. In the first study, we introduce a new transtheoretical taxonomy of therapeutic interventions and examine its application across a sample of Czech psychotherapists. The resulting distribution of intervention categories provides an empirical foundation for automation. In the second study, we assess the performance of different language-modeling approaches for predicting these intervention categories from session transcripts, comparing fine-tuned masked language models with large pretrained language models evaluated in zero- and few-shot settings. Model accuracy, contextual sensitivity, and efficiency are analyzed to identify the most effective strategy for automatic intervention recognition. Together, the studies demonstrate the feasibility of applying modern language modeling techniques to the analysis of psychotherapy process, highlighting practical implications for research, training, and feedback systems.