Predicting Client Emotions and Therapist Interventions in Psychotherapy Dialogues

Natural Language Processing (NLP) can advance psychotherapy research by scaling up therapy dialogue analysis as well as by allowing researchers to examine client-therapist interactions in detail. Previous studies have mainly either explored the clients' behavior or the therapists' intervention in dialogues. Yet, modelling conversations from both dialogue participants is crucial to understanding the therapeutic interaction. This study explores speaker contribution-based dialogue acts at the utterance-level; i.e, the therapist-Intervention Prediction (IP) and the client-Emotion Recognition (ER) in psychotherapy using a pan-theoretical schema. We perform experiments with fine-tuned language models and light-weight adapter solutions on a Hebrew dataset. We deploy the results from our ER model predictions in investigating the coherence between client self-reports on emotion and the utterance-level emotions. Our best adapters achieved on-par performance with fully fine-tuned models, at 0.64 and 0.66 micro F1 for IP and ER, respectively. In addition, our analysis identifies ambiguities within categorical clinical coding, which can be used to fine-tune the coding schema. Finally, our results indicate a positive correlation between client self-reports and utterance-level emotions.