💬 NLP

Graph2Counsel: Clinically Grounded Synthetic Counseling Dialogue Generation from Client Psychological Graphs

The paper introduces Graph2Counsel, a framework that generates high-quality, clinically grounded synthetic counseling dialogues by leveraging structured Client Psychological Graphs to overcome data scarcity and privacy limitations, resulting in a dataset that significantly improves the performance of fine-tuned language models on counseling benchmarks.

Original authors: Aishik Mandal, Hiba Arnaout, Clarissa W. Ong, Juliet Bockhorst, Kate Sheehan, Rachael Moldow, Tanmoy Chakraborty, Iryna Gurevych

Published 2026-04-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Aishik Mandal, Hiba Arnaout, Clarissa W. Ong, Juliet Bockhorst, Kate Sheehan, Rachael Moldow, Tanmoy Chakraborty, Iryna Gurevych

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to teach a robot how to be a compassionate, professional therapist. You can't just show it a textbook; you need to show it real conversations. But here's the problem: real therapy sessions are like gold dust. They are incredibly valuable, but they are also locked away behind heavy vaults of privacy laws. You can't just hand them to an AI because that would violate people's trust and safety.

So, researchers had to build a "fake" therapy dataset. But previous attempts were like trying to teach someone to drive by showing them a picture of a car. They had the look of a conversation, but they missed the engine—the deep, complex reasons why a person feels the way they do.

Enter Graph2Counsel. Think of this not as a writer, but as a master architect building a realistic simulation of the human mind.

The Blueprint: The "Mind Map" (Client Psychological Graph)

Most AI training data is just a list of symptoms: "Client is sad," "Client is anxious." It's flat. It's like a grocery list.

Graph2Counsel uses something called a Client Psychological Graph (CPG). Imagine this as a dynamic, living subway map of a person's mind.

The Stations (Nodes): These are feelings or thoughts (e.g., "Fear of Judgment," "Overthinking").
The Tracks (Edges): These show how the stations connect. One track might say, "Fear of Judgment triggers Overthinking." Another might say, "Mindfulness calms Fear of Judgment."

This map doesn't just list problems; it shows the cause-and-effect relationships inside a person's head. It captures the flow of a human experience, not just a snapshot.

The Construction Site: Building the Dialogue

Once the researchers have this "Mind Map," they use it to build synthetic therapy sessions. Here is how they do it, using a few clever tricks:

The Actors (Client Profiles): They take one "Mind Map" and dress it up in different costumes. One map might become "Elena, a 28-year-old graphic designer who worries about her boss," and the same map might become "Raj, a 35-year-old accountant who feels misunderstood by his wife." The underlying psychology is the same, but the story changes.
The Director (Counselor Strategies): They don't just let the AI chat randomly. They give the AI a script of real techniques used by human therapists (like "Empathy Building" or "Reframing"). It's like giving the AI a director who whispers, "Now, try to validate their feelings," or "Now, ask a gentle question."
The Rehearsal (Multi-Agent Feedback): Before the final performance, the AI acts out the scene. Then, a second "critic AI" watches and says, "Hey, that counselor sounded too robotic," or "The client agreed too quickly; real people are usually more confused." The first AI then rewrites the scene to fix it.

The Result: A New Benchmark

The result is 760 therapy sessions that feel surprisingly real.

The Test: When human experts (licensed therapists) reviewed these conversations, they rated them higher than any previous fake dataset. They said the conversations flowed better, the counselor sounded more competent, and the client felt more authentic.
The Safety Check: Crucially, these sessions were very safe. The AI didn't give dangerous medical advice or hurtful comments, because it was grounded in the structured "Mind Map" and real therapist strategies.

Why This Matters

Think of this like flight simulators for pilots.

Old way: You tried to learn to fly by reading a manual or watching a video of a plane. (This is what previous AI datasets did).
Graph2Counsel way: You are in a simulator that understands the physics of wind, the mechanics of the engine, and the stress of the pilot. You can crash the plane a thousand times in the simulator without anyone getting hurt.

By training AI on these high-quality, "Mind Map" grounded simulations, we can create mental health tools that are:

Safer: Less likely to give bad advice.
Smarter: Better at understanding the why behind a person's feelings.
More Available: Helping people who can't afford or access a human therapist right now.

In short, Graph2Counsel is building a safe, realistic training ground so that AI can learn to be a helpful, empathetic companion for our mental health, without ever needing to peek into a real person's private diary.

1. Problem Statement

The deployment of Large Language Models (LLMs) for mental health counseling is hindered by a critical lack of high-quality, real-world training data due to strict privacy regulations and ethical constraints. Existing synthetic datasets often rely on unstructured or semi-structured inputs (e.g., static client profiles, symptom lists, or questionnaires) that fail to capture the functional dependencies between a client's cognitive, emotional, and behavioral states. This omission leads to psychologically inconsistent interactions, hallucinated advice, and a lack of clinical realism in generated dialogues. The core challenge is to generate synthetic counseling data that is not only diverse but also grounded in the complex, dynamic causal relationships inherent in human psychology.

2. Methodology: Graph2Counsel Framework

The authors propose Graph2Counsel, a framework that generates synthetic counseling sessions grounded in Client Psychological Graphs (CPGs).

A. Core Representation: Client Psychological Graphs (CPGs)

Instead of using raw text transcripts, the framework utilizes CPGs, which are structured, directed graphs derived from real therapy sessions.

Nodes: Represent psychological processes (e.g., "fear of judgment," "overthinking," "mindfulness ability").
Edges: Represent functional relationships, specifically excitatory (e.g., fear triggers overthinking) or inhibitory (e.g., mindfulness reduces fear) influences.
Advantage: CPGs abstract away sensitive personal identifiers while preserving the causal dynamics of the client's internal state, allowing for scalable and privacy-preserving data generation.

B. Data Generation Pipeline

The framework operates in two main stages:

Profile Generation: From a single CPG, the system generates diverse Client Profiles. These are semi-structured summaries including demographics, presenting problems, and history. Crucially, one CPG can generate multiple unique profiles (e.g., different ages, genders, occupations) that share the same underlying psychological dynamics, enabling dataset expansion without compromising the structural integrity of the psychological model.
Dialogue Generation: The system synthesizes multi-turn counseling sessions using the CPG, the generated profile, and Counselor Strategies (interventions like reframing or empathy building) extracted from real sessions.

C. Prompting Strategies

The authors explored and compared four prompting techniques to optimize dialogue quality:

Base: Standard generation without specific counseling guidance.
Guided Counseling (GC): Conditioning the LLM on high-level strategies extracted from real transcripts to ensure clinically appropriate interventions.
GC + Chain-of-Thought (CoT): Requiring the model to generate intermediate reasoning steps grounded in the CPG and profile before producing an utterance.
GC + Multi-Agent (GC+MA): An iterative workflow where one agent generates a session and a second agent critiques it against guidelines, followed by a regeneration step.

Optimal Configuration: Through ablation studies, the authors determined that Guided Counseling (GC) combined with CPG + Profile inputs yielded the best balance of quality, cost, and dialogue length. While CoT and Multi-Agent approaches showed marginal score improvements, they incurred significantly higher computational costs and produced shorter or less natural dialogues.

3. Key Contributions

Graph2Counsel Framework: A novel approach for generating synthetic counseling data grounded in CPGs, capturing the interplay between thoughts, emotions, and behaviors.
Dataset Construction: Creation of a dataset comprising 760 synthetic counseling sessions derived from 76 unique CPGs. The dataset covers 29 distinct therapy modalities (e.g., CBT, DBT, ACT) and diverse client profiles.
Comprehensive Evaluation:
- Expert Evaluation: Conducted by four licensed clinicians using the Cognitive Therapy Rating Scale (CTRS) and Working Alliance Inventory (WAI).
- Downstream Utility: Fine-tuning an open-source model (Llama3-8B-Instruct) on the dataset and evaluating it on standard benchmarks (CounselingBench and CounselBench).
- Faithfulness: Quantitative metrics ensuring the generated dialogues accurately reflect the input CPGs and profiles.
Open Source: Public release of code, data, and prompts to facilitate further research in mental health AI.

4. Results

Expert Evaluation: Graph2Counsel outperformed state-of-the-art baselines (CACTUS, MAGneT, SQPsychConv) across all metrics:
- Specificity, Counselor Competence, Authenticity, Conversational Flow, and Safety.
- It achieved the lowest rate of unsafe utterances (0.5%).
- Inter-annotator agreement was substantial (Krippendorff's $\alpha$ = 0.70).
Faithfulness: The generated sessions showed high fidelity to inputs:
- CPG Faithfulness: 91% of psychological processes in the graph were reflected in the dialogue.
- Profile Faithfulness: 99% of sessions contained no contradictions with the client profile.
Downstream Performance:
- Fine-tuning Llama3-G2C (the model trained on Graph2Counsel) significantly improved performance on CounselingBench (multiple-choice clinical competency questions) and CounselBench (open-ended response quality).
- Llama3-G2C achieved the highest scores in empathy, specificity, and factual consistency, while maintaining the lowest rates of providing inappropriate medical advice or toxic responses.
- In adversarial testing (CounselBench-Adv), the model demonstrated robustness, though it showed some vulnerability in "Symptom" probing compared to other baselines, highlighting a trade-off between psychological grounding and safety in specific failure modes.

5. Significance and Impact

Bridging the Data Gap: Graph2Counsel offers a scalable solution to the scarcity of mental health data by leveraging structured psychological graphs rather than relying solely on scarce, privacy-sensitive transcripts.
Clinical Realism: By modeling the causal dependencies between cognitive and emotional states, the framework produces dialogues that are psychologically coherent, moving beyond superficial symptom matching to simulate genuine therapeutic reasoning.
Safety and Ethics: The framework prioritizes safety, with rigorous expert evaluation ensuring that generated content adheres to therapeutic norms and minimizes harmful advice.
Future Directions: The authors acknowledge limitations regarding scalability (current reliance on a limited set of CPGs) and the need for more diverse demographic representation. However, the framework provides a robust meta-framework for integrating process-based therapy into AI systems, paving the way for safer, more effective, and clinically grounded mental health support tools.