Introduction
The global prevalence of depression presents a significant public health challenge, with rates in China reaching approximately 7%. Social media platforms, particularly "treehole" threads on Weibo, have emerged as critical spaces for individuals to express emotional distress and suicidal ideation. However, a substantial "counselor gap" exists; the demand for psychological intervention far outstrips the supply of trained professionals. Consequently, many interventions are led by well-meaning but untrained volunteers. These non-professionals, while motivated, risk causing inadvertent harm through insensitive language or a failure to recognize critical psychological triggers. This paper introduces a multifaceted decision-support system powered by GPT-3.5, designed to act as a "safety net" for these volunteers, providing real-time, expert-level guidance.
System Architecture and Methodology
The LLM-Counselors Support System employs a human-in-the-loop workflow where the AI acts as a supervisor rather than a replacement for the human counselor. The system processes counselor-client interactions through several specialized layers to ensure safety and efficacy.
Technical Components
- Privacy Information Filtering: To protect user anonymity, the system utilizes regular expressions (regex) to mask personally identifiable information (PII) such as names and contact details before data reaches the LLM.
- Structured Prompt Engineering: The system's core logic is governed by a sophisticated prompt structure covering seven dimensions: Role Definition (psychological expert), Task Definition, Role Boundaries, Contextual Requirements, Error Identification, Resource Provision, and Cognitive Distortion Classification.
- Detoxification Pipeline: To mitigate the risk of generating harmful content, a similarity-based filter was implemented. Using the COLD (Chinese Offensive Language Dataset), the authors created a vector database. Generated responses are compared against this database using L2 distance; if a response is deemed too close to offensive content, it undergoes iterative refinement.
- Cognitive Distortion Identification: The system is specifically tuned to detect "thinking traps" like over-attribution and overgeneralization, providing counselors with a deeper understanding of the client's psychological state.
Evaluation and Results
The system was rigorously evaluated by ten professional psychotherapists across five key dimensions. The results demonstrate high alignment between the LLM's outputs and expert judgment.
Key performance metrics include:
- Accuracy of Problem Analysis: Achieved a 93% "Yes" rating in single-round interactions and 97.5% in multi-round scenarios.
- Cognitive Distortion Identification: Successfully identified distortions in 91.5% of cases.
- Counselor Behavior Assessment: Detected unprofessional behavior in 94.5% of single-round interactions.
- Inter-rater Reliability: The system achieved a Krippendorff’s Alpha of 0.95 to 0.98, indicating exceptionally high consistency among expert evaluators.
Discussion and Limitations
The system effectively transitions volunteers from judgmental or insensitive language toward supportive, exploratory dialogue through its "Recommended Response Scripts" (Reply+). By evaluating the counselor's own behavior, the LLM serves as a "meta-counselor," pointing out instances where a volunteer might be overly assertive or irresponsible.
Despite these successes, several limitations remain. Experts noted that LLM-generated responses can occasionally feel "sterile" or "insincere," lacking the warmth of genuine human empathy. Additionally, the system sometimes struggles with the nuances of internet slang and emoticons common in social media discourse. There is also a noted risk of "cognitive stagnation," where volunteers might become overly dependent on the system's suggestions rather than developing their own clinical intuition. Future work will focus on refining the model's handling of complex emotional expressions and conducting broader real-world testing with diverse volunteer groups.
Summary updated