Consistent Client Simulation for Motivational Interviewing-based Counseling

TL;DR

This work addresses the challenge of training and evaluating Motivational Interviewing counselors using consistent client simulations. It introduces a four-module, state-transition driven framework—comprising state transition, action selection, information selection, and response generation—that enforces alignment between client utterances and rich client profiles (motivation, beliefs, preferred change plans, and receptivity) and MI dynamics. The framework leverages real-world MI data from the AnnoMI dataset and GPT-4-based annotations to derive knowledge for the modules, and it is evaluated against multiple baselines through automatic and expert assessments. Results show superior profile-consistency and receptivity alignment, as well as realistic action distributions and session lengths compared to baselines, though real human sessions still outperform simulated ones. The approach holds promise for scalable training and evaluation of MI counselors and can be extended to additional counseling modalities as more real-world data become available.

Abstract

Simulating human clients in mental health counseling is crucial for training and evaluating counselors (both human or simulated) in a scalable manner. Nevertheless, past research on client simulation did not focus on complex conversation tasks such as mental health counseling. In these tasks, the challenge is to ensure that the client's actions (i.e., interactions with the counselor) are consistent with with its stipulated profiles and negative behavior settings. In this paper, we propose a novel framework that supports consistent client simulation for mental health counseling. Our framework tracks the mental state of a simulated client, controls its state transitions, and generates for each state behaviors consistent with the client's motivation, beliefs, preferred plan to change, and receptivity. By varying the client profile and receptivity, we demonstrate that consistent simulated clients for different counseling scenarios can be effectively created. Both our automatic and expert evaluations on the generated counseling sessions also show that our client simulation method achieves higher consistency than previous methods.

Figures (5)

• Figure 1: Types of inconsistency in existing client simulation approaches (Left: simulated client, Right: counselor): client response inconsistent with the (a) client's motivation to change, (b) beliefs, (c) preferred plans to change, and (d) receptivity.
• Figure 2: Proposed Client Simulation Framework.
• Figure 3: Distribution of turn count (Count) for various clients. The length of AnnoMI counseling sessions is diverse and generally longer, while simulated sessions tend to have fewer than 50 turns.
• Figure 4: The distribution of receptivity and the relation between receptivity and sustain ratio and precontemplation.
• Figure 5: Proportions of actions for different receptivity scores in different states. Actions negatively associated with receptivity are represented in dashed lines.