DREAM-Meta learning for tailored hierarchical dialogue management: application to motivational interview

Area of Science:

• Artificial Intelligence
• Human-Computer Interaction
• Computational Linguistics

Background:

• Large Language Models (LLMs) offer rich interactions but lack control and structured adaptation for complex, multi-stage tasks.
• Current dialogue systems often fail to address both user personalization and dynamic goal management simultaneously.
• Nuanced, goal-driven conversations require adaptation to diverse user needs and evolving interaction phases.

Purpose of the Study:

• To propose a hybrid dialogue framework combining LLMs with a hierarchical dialogue manager.
• To enable two levels of adaptation: personalization for user profiles and dynamic management of interaction sub-goals.
• To test the framework's efficacy in a challenging domain like Motivational Interviewing (MI).

Main Methods:

• Integration of a Large Language Model (LLM) with a hierarchical dialogue manager.
• Utilization of hierarchical reinforcement learning to model multi-phase dialogue progression (e.g., MI).
• Application of meta-learning for rapid adaptation across diverse user profiles.
• Evaluation through simulated users and human participant experiments.

Main Results:

• The hybrid system demonstrated more engaging interactions compared to a standard LLM baseline.
• The framework successfully managed both conversational structure and user-specific personalization.
• Effectiveness shown in the complex, adaptive context of Motivational Interviewing.

Conclusions:

• Hybrid dialogue architectures are valuable for achieving sophisticated, two-level adaptation in conversational AI.
• The proposed framework enhances dialogue system capabilities for goal-driven and personalized interactions.
• This approach advances the development of more effective and adaptable AI-powered communication tools.