Rethinking the Alignment of Psychotherapy Dialogue Generation with Motivational Interviewing Strategies

TL;DR

The paper tackles the challenge of aligning LLM-based psychotherapy dialogue with Motivational Interviewing (MI) principles to improve safety, controllability, and explainability. It introduces a strategy-aligned prompting framework in which the model first predicts an MI strategy (MI skill code) and then generates the therapist's utterance in strict accordance with that strategy, leveraging Chain-of-Thought-like internal reasoning. The authors validate the approach with automatic metrics and comprehensive human evaluations on two MI datasets (AnnoMI and BiMISC), comparing multiple open-source LLMs and GPT-4, and show that strategy-aligned prompts generally improve MI adherence and perceived quality while maintaining reasonable flexibility. The work demonstrates the potential for safe, explainable MI dialogue generation with LLMs and outlines practical considerations, limitations, and directions for future real-world deployment and ethical evaluation in psychotherapy.

Abstract

Recent advancements in large language models (LLMs) have shown promise in generating psychotherapeutic dialogues, particularly in the context of motivational interviewing (MI). However, the inherent lack of transparency in LLM outputs presents significant challenges given the sensitive nature of psychotherapy. Applying MI strategies, a set of MI skills, to generate more controllable therapeutic-adherent conversations with explainability provides a possible solution. In this work, we explore the alignment of LLMs with MI strategies by first prompting the LLMs to predict the appropriate strategies as reasoning and then utilizing these strategies to guide the subsequent dialogue generation. We seek to investigate whether such alignment leads to more controllable and explainable generations. Multiple experiments including automatic and human evaluations are conducted to validate the effectiveness of MI strategies in aligning psychotherapy dialogue generation. Our findings demonstrate the potential of LLMs in producing strategically aligned dialogues and suggest directions for practical applications in psychotherapeutic settings.

Figures (6)

• Figure 1: Visual summary of the research questions and corresponding experimental evaluations did in this work.
• Figure 2: The "strategy-aligned" dialogue generation with the strategy prediction as step 1 and the subsequent therapist's utterances generation as step 2 in the context of Motivational Interviewing.
• Figure 3: Comparison of attention score distributions from LLM (i.e., Flan-T5 in \ref{['appendix: llms']} with Encoder-Decoder architecture, last layer, and averaged across all heads) for dialogue generation, with ("strategy-aligned") and without ("standard") the MI strategy. The input conversational context and strategy match the 'Case Study' example for consistency. Attention to input tokens is aggregated into three prompting components for better comparison.
• Figure 4: Experts evaluation on two datasets based on assessing criteria ("EC1-EC5") in Table \ref{['tab:ec_items']}. It assesses the alignment between MI strategy and utterances generated by "Standard" and "Strategy-aligned" prompts. The y-axis denotes the average ratings ranging from 1 (Strongly disagree) to 5 (Strongly agree). (**$p$<.01, *$p$<.05, "ns" $p$<.1)
• Figure 5: Experts assess the MI strategy prediction by GPT-4 using criteria (EC6): "MI strategy aligns with dialogue context and MI principles". (**$p$<.01, *$p$<.05)