Abstract:

In an increasingly complex world, more and more people are seeking online advice from professional counselors for psychological or social issues. The spectrum of cases ranges from relationship problems and sexuality to addiction, loss, and grief and therefore encompasses all problems of people in different life situations and circumstances. Training competent counselors is a challenging task and requires not only teaching specific methods but also simulating realistic counseling scenarios through role-playing exercises. Role-plays with humans, while valuable, are often limited by the availability of supervisors and role-play partners, restricting opportunities for independent practice.

The article gives an overview of VirCo (Virtual Client for Online Counseling), an innovative learning platform that utilizes an open-source large language model (LLM) at its core to simulate clients with psycho-social problems. The system's dataset comprises detailed persona descriptions and transcripts from actual role-play sessions, contributing to the authenticity of the training experience. Thus, learners can independently engage with various client personas in realistic conversations and acquire counseling experiences in a safe and controlled environment without direct trainer supervision.

Building such a virtual client is challenging because, on the one hand, it needs to stick to a given case description, and on the other hand, it must reply appropriately to any counselor question or recommendation that may come up during the conversation. Earlier approaches to counseling chatbots mostly failed to cope with this kind of dynamic conversation flow. The situation changed fundamentally with the availability of large language models (LLMs) like ChatGPT or Llama and its variants. Our experiments show that these models provide an excellent basis for conversational sparring partners.

The chatbot is embedded in a learning platform which provides functionality to create courses, define compulsory assignments, and implements multiple feedback mechanisms. The platform's coin-based incentive system encourages peer feedback, fostering a collaborative learning environment. Feedback can also be generated automatically by a large language model. These features not only enhance the learning experience but also facilitate continuous improvement and skill development for trainees.

In the article, we explain how to customize large language models to simulate different personas, introduce the different feedback mechanisms, and present first results from evaluations of actual training sessions in the course "Introduction to Online Counseling", which is part of the Bachelor's program in Social Work at our university.

In summary, the Virtual Client represents a novel application of AI technology for the education of online counselors. By leveraging the power of LLMs, it offers a scalable and accessible solution for trainees to gain practical experience through simulated client interactions, ultimately contributing to the development of a skilled and well-prepared workforce in the rapidly growing field of online counseling.

Keywords:

Education, Role-Play, Counseling, Generative AI.