[Paper Review] Enhancing Psychological Counseling with Large Language Model: A Multifaceted Decision-Support System for Non-Professionals

Guanghui Fu, Qing Zhao+10 more|arXiv (Cornell University)|Aug 29, 2023

Mental Health via Writing17 citations

TL;DR

The paper proposes a large language model–based counselor support system to aid non-professionals in online psychological interventions, validating it with expert evaluation across multiple dimensions.

ABSTRACT

In the contemporary landscape of social media, an alarming number of users express negative emotions, some of which manifest as strong suicidal intentions. This situation underscores a profound need for trained psychological counselors who can enact effective mental interventions. However, the development of these professionals is often an imperative but time-consuming task. Consequently, the mobilization of non-professionals or volunteers in this capacity emerges as a pressing concern. Leveraging the capabilities of artificial intelligence, and in particular, the recent advances in large language models, offers a viable solution to this challenge. This paper introduces a novel model constructed on the foundation of large language models to fully assist non-professionals in providing psychological interventions on online user discourses. This framework makes it plausible to harness the power of non-professional counselors in a meaningful way. A comprehensive study was conducted involving ten professional psychological counselors of varying expertise, evaluating the system across five critical dimensions. The findings affirm that our system is capable of analyzing patients' issues with relative accuracy and proffering professional-level strategies recommendations, thereby enhancing support for non-professionals. This research serves as a compelling validation of the application of large language models in the field of psychology and lays the groundwork for a new paradigm of community-based mental health support.

Motivation & Objective

• Address the shortage of trained心理 counselors by enabling non-professionals to provide online psychological support using LLMs.
• Develop a safe, privacy-conscious system that assists non-professionals with analysis and response strategies.
• Validate system effectiveness and safety through expert evaluation across multiple rounds of conversation.

Proposed method

• Construct an LLM-counselor support workflow where counselors input initial responses and client comments, and the system provides an enhanced Reply+ for counselor review.
• Implement privacy information filtering using regular expressions to mask identifiable data prior to processing.
• Design a structured prompt that defines the AI role, tasks, boundaries, context, error handling, resources, cognitive distortion identification, and input/output formats.
• Use a detoxification pipeline that compares generated content against an offensive dataset via embeddings and a vector database to ensure non-offensive outputs (threshold α = 0.2).
• Generate reports from the LLM outputs and have them evaluated by human counselors to assess accuracy and safety across five dimensions.
• Discuss system design considerations including expert oversight, ethical constraints, and potential limitations for real-world deployment.

Experimental results

Research questions

• RQ1 Can the LLM-counselor system analyze a client’s problems and cognitive distortions with accuracy comparable to professional standards?
• RQ2 Are the generated verbal strategies appropriate and effective for non-professional counselors?
• RQ3 Can the system provide useful suggestions for next steps while maintaining safety and privacy?
• RQ4 How reliable are expert evaluations of system-generated reports across single-round and multi-round dialogues?
• RQ5 What are the key limitations and safety concerns for deploying such a system in social media contexts?

Key findings

• Experts rated the system as achieving high accuracy in patient problem analysis (≈93% in single-round; ≈97.5% in multi-round).
• Cognitive distortion analysis accuracy was high (≈91.5% in single-round; ≈95% in multi-round).
• Counselor behavior analysis accuracy was high in single-round (≈94.5%), but slightly lower in multi-round contexts (≈85%).
• Appropriateness of verbal strategies was about 84% for single-round and 78.75% for multi-round conversations, with notable uncertainty.
• Effective next-step suggestions were provided with ≈87% accuracy in single-round and ≈82.5% in multi-round evaluations.
• Krippendorff’s Alpha indicated high inter-rater reliability (0.98 for single-round, 0.95 for multi-round).