Towards Automated Counselling Decision-Making: Remarks on Therapist Action Forecasting on the AnnoMI Dataset

Substantial progress has been made in recent years on natural language processing approaches to counselling conversation analysis. However, few studies have investigated therapist action forecasting, which aims to suggest dialogue actions that the therapist can take in the next turn, partly due to generally limited access to counselling dialogue data resulting from privacy-related constraints. In this work, we leverage a recently released public dataset of therapy conversations and experiment with a range of natural language processing techniques to approach the task of therapist action forecasting with language models. We probe various factors that could impact model performance, including data augmentation, dialogue context length, incorporating therapist/client utterance labels in the input, and contrasting high- and low-quality counselling dialogues. With our findings, we hope to provide insights on this task and inspire future efforts in counselling dialogue analysis.