StoryMI: Multi-Agent Framework for Therapeutic Dialogue Generation
A new framework called StoryMI has been developed by researchers to facilitate controllable motivational interviewing (MI) dialogue generation through a multi-LLM agent approach. This system enhances client profiles based on questionnaires, transforming them into situational narratives for context. During multi-turn conversations, both therapist and client agents produce MI-coded responses, with an interaction agent selecting MI codes to manage the dialogue strategies. The evaluation process consists of two levels, incorporating lexical metrics and MI-focused assessments of overarching counseling strategies, as well as evaluations from both LLM and human experts. A dataset comprising 6,000 simulated MI dialogues was created, based on 1,000 questionnaire-story pairs that represent 12 MI codes.
Key facts
- StoryMI is a multi-LLM agent framework for controllable MI dialogue generation.
- Client profiles are expanded into situational stories for narrative context.
- Therapist and client agents generate MI-coded utterances.
- An interaction agent dynamically coordinates exchanges to control MI strategies.
- Evaluation includes lexical metrics and MI-specific measures.
- LLM-as-judge and human expert assessments are used.
- Dataset: 6,000 simulated MI dialogues from 1,000 questionnaire-story pairs.
- Covers 12 MI codes.
Entities
Institutions
- arXiv