Graph2Counsel: Clinically Grounded Synthetic Counseling Dialogue Generation from Clinical Graphs¶
Conference: EMNLP 2026
Paper: ACL Anthology
Code: GitHub
Area: Medical NLP
Keywords: Counseling Dialogue Generation, Client Psychological Graphs (CPGs), Process-Based Therapy, Clinical Strategies, LLM Fine-Tuning
TL;DR¶
Graph2Counsel introduces a clinically grounded synthetic counseling dialogue generation framework anchored in Client Psychological Graphs (CPGs) and extracted therapist strategies, producing realistic multi-turn sessions that outperform existing synthetic benchmarks across expert evaluations and downstream counseling tasks.
Background & Motivation¶
Global demand for mental health support has surged dramatically, but severe clinician shortages, high financial costs, and pervasive social stigma prevent millions from accessing timely care. Large Language Models (LLMs) offer promising avenues for scalable and non-judgmental assistance; however, deploying general-purpose LLMs directly in therapy-critical settings poses severe risks, including hallucinated guidance, inappropriate interventions, and exacerbation of client distress. Mitigating these hazards demands high-quality training conversations, but acquiring real-world therapy transcripts is heavily constrained by strict confidentiality protocols and ethical barriers, while automated pseudonymization and manual de-identification remain vulnerable to privacy leakage.
Existing synthetic counseling datasets predominantly rely on flat or unstructured inputs to simulate dialogues. While initial benchmarks relied on single-turn QA pairs, subsequent datasets introduced multi-turn conversations conditioned on static personas, symptom checklists, or diagnostic questionnaire scores. Nevertheless, these inputs capture only disconnected snapshots of a client's condition and fail to account for the functional dependencies linking cognitive, emotional, and behavioral processes. In actual therapy, identifying that a client experiences both "overthinking" and "fear of judgment" is insufficient; models must reflect how fear of judgment actively excites rumination, or how mindfulness practices mitigate negative affect, in order to maintain therapeutic coherence.
To overcome the fragility of unstructured generation, clinical research offers a principled structural representation: Client Psychological Graphs (CPGs). CPGs capture mental functioning as directed symptom networks where nodes represent latent psychological mechanisms and edges denote excitatory or inhibitory relations. Crucially, CPGs abstract away personal identifying details while preserving functional pathology dynamics. Core idea: Ground synthetic counseling dialogue generation in causally structured Client Psychological Graphs (CPGs) and extracted therapist strategies, deriving diverse client profiles from the graph and generating multi-turn sessions under clinical guidelines to balance authenticity, clinical competence, and privacy.
Method¶
Overall Architecture¶
Graph2Counsel translates the latent clinical dynamics captured in real therapy sessions into diverse, multi-turn, and privacy-preserving counseling dialogues. The overarching generation pipeline comprises four primary stages: first, extracting CPG networks and therapist strategies from de-identified clinical transcripts; second, sampling diverse client intake profiles conditioned on each CPG to ground abstract nodes in concrete life circumstances without exposing personal identifiers; third, prompting an LLM with the joint conditioning of CPG edges, client intake forms, and extracted therapist strategies under expert clinical guidelines; and fourth, assessing the generated sessions across clinical scales and fine-tuning an open-source model via QLoRA for downstream task validation.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["De-identified Clinical Transcripts<br/>Real Therapy Sessions"] --> B["Stage 1: Graph & Strategy Extraction<br/>CPG Topology + Interventions"]
B --> C["Stage 2: Client Profile Diversification<br/>10 Realistic Personas per Graph"]
C --> D["Stage 3: Joint Guided Dialogue Generation<br/>CPG + Profile + Strategy Prompting"]
D --> E["Stage 4: Clinical Benchmarking & Model Distillation<br/>CTRS / WAI / Expert Review / QLoRA Tuning"]
Key Designs¶
1. Graph & Strategy Extraction: Decoupling Personal Identifiers from Clinical Dynamics
Directly generating synthetic dialogues from raw transcripts risks leaking sensitive details, whereas generating from basic personas yields shallow conversations devoid of clinical reasoning. To address this tension, this design establishes a dual extraction pipeline: first, it identifies latent cognitive, emotional, and behavioral processes from transcripts, clusters them into themes, and infers directed excitatory or inhibitory relations to construct Client Psychological Graphs (averaging 10.84 nodes and 57.88 edges per graph across 76 real sessions). Concurrently, a locally hosted Llama3.1-70B-Instruct extracts 257 distinct counselor intervention strategies (e.g., reframing, thought experiments, empathy building, evidence-based questioning) alongside their supporting evidence utterances, mapping them across evidence-based therapy modalities such as CBT, DBT, ACT, and person-centered therapy under a process-based therapy framework. By separating structural pathology from patient identity, the method yields clean, reusable therapeutic building blocks.
2. Client Profile Diversification: Grounding Abstract Networks into Rich Realities
If a single psychological graph can only produce a single session, the resulting synthetic corpus lacks demographic breadth and situational variability. This design introduces an automated intake form generation protocol that conditions an LLM on the CPG while strictly forbidding clinical jargon, forcing the model to articulate problems from the client's lay perspective. For each CPG, the framework generates 10 distinct client profiles that vary across age, gender, occupation, family background, and life stressors while sharing the identical underlying graph structure. For example, the same node representing "difficulty adjusting to new roles" is realized as a 37-year-old chef opening a new restaurant in one profile, and as a 50-year-old widowed father navigating household and career responsibilities in another. This diversification multiplies 76 seed graphs into 760 full-scale client profiles with 97% rated realistic by clinicians.
3. Joint Guided Dialogue Generation: Conditioning on CPG Topology and Therapist Strategies
Standard LLM generation frequently suffers from behavioral artifacts such as clients agreeing too readily, displaying unrealistic clinical self-insight, or counselors abruptly jumping across topics. To enforce therapeutic fidelity, this design establishes Guided Counseling (GC) prompting, which embeds global conversational constraints, client-specific guidelines, counselor guidelines, and explicit negative examples (common pitfalls) co-designed with clinicians. The generation prompt supplies the CPG edge list, client profile, and relevant counselor strategies, instructing the model to produce rich interactions of at least 40 turns. Ablations across input formats and prompting techniques demonstrated that while Chain-of-Thought (GC+CoT) slightly improves automated metrics, it prematurely terminates conversations (30 turns vs. 40 turns), and Multi-Agent feedback (GC+MA) doubles generation cost without significant perceptual gain. Consequently, GC with joint CPG+Profile input was adopted, delivering 91% CPG faithfulness and 99% profile consistency while sustaining deep, natural pacing.
Loss & Training¶
To measure the downstream efficacy of the synthesized corpus, the authors fine-tune an open-source Llama3-8B-Instruct model on Graph2Counsel using QLoRA (denoted as Llama3-G2C). The optimization objective follows the standard causal language modeling loss over counselor responses:
where \(x_{\mathrm{context}}\) encapsulates the preceding dialogue turns and \(y_t\) represents the ground-truth counselor response tokens. The 760 sessions are partitioned into 14,597 training instances of (dialogue_history, counselor_response). Fine-tuning is conducted with LoRA rank \(r = 64\) and \(\alpha = 16\), learning rate \(1 \times 10^{-5}\), dropout \(0.1\), and batch size \(4\) across four A100 80GB GPUs over 3 epochs using DeepSpeed and the Hugging Face Trainer.
Key Experimental Results¶
Main Results¶
Four licensed clinical psychologists conducted an extensive blinded evaluation over 100 semantically aligned client issues across Graph2Counsel and three state-of-the-art multi-turn synthetic datasets (CACTUS, MAGneT, and SQPsychConv). Metrics reflect average rank (1โ4, lower is better) and percentage of unsafe sessions:
| Dataset | Specificity (โ) | Counselor Competence (โ) | Authenticity (โ) | Conversational Flow (โ) | Unsafe Sessions (%) (โ) |
|---|---|---|---|---|---|
| CACTUS (Lee et al., 2024) | 2.41 | 2.39 | 2.22 | 2.00 | 3.0 |
| MAGneT (Mandal et al., 2026) | 3.97 | 3.98 | 3.99 | 3.99 | 1.0 |
| SQPsychConv (Vu et al., 2025) | 1.84 | 1.97 | 2.37 | 2.53 | 1.0 |
| Graph2Counsel (Ours) | 1.79 | 1.67 | 1.43 | 1.48 | 0.5 |
Inter-annotator agreement achieved Krippendorff's \(\alpha = 0.70\) (and 91% consensus on safety), demonstrating strong clinician consensus. Graph2Counsel secured the top rank across all dimensions, with substantial margins in Authenticity (1.43) and Flow (1.48).
Downstream evaluation on CounselBench-Eval assessed the fine-tuned Llama3-8B-Instruct models using GPT-4o as judge:
| Model | Overall Quality (1โ5) (โ) | Empathy (1โ5) (โ) | Specificity (1โ5) (โ) | Medical Advice (%) (โ) | Factual Consistency (1โ4) (โ) | Toxicity (1โ5) (โ) |
|---|---|---|---|---|---|---|
| CAMEL (CACTUS) | 3.90*** | 3.94*** | 3.91*** | 4.0 | 3.97 | 1.00 |
| Llama3-MAG | 4.12* | 4.09** | 4.12** | 8.0 | 3.98 | 1.00 |
| Llama3-SQP | 4.07*** | 4.06*** | 4.08*** | 7.0 | 3.98 | 1.00 |
| Llama3-G2C (Ours) | 4.29 | 4.26 | 4.30 | 4.0 | 3.98 | 1.00 |
| \(\delta(\%)\) vs Best Baseline | +4.25% | +4.25% | +4.50% | 0.00% | 0.00% | 0.00% |
Note: * \(p < 0.05\), ** \(p < 0.01\), *** \(p < 0.001\) from paired t-tests.
Ablation Study¶
Automated evaluation using GPT-4o on the Cognitive Therapy Rating Scale (CTRS, 0โ6) and Working Alliance Inventory (WAI, 1โ7) across 15 generation configurations:
| Prompting Technique | Input Type | CTRS General (U/I/C) | CTRS CBT Skills (D/F/S) | WAI Task (โ) | WAI Goal (โ) | WAI Bond (โ) | Avg. Turns |
|---|---|---|---|---|---|---|---|
| Base | CPG | 4.68 / 5.97 / 5.18 | 4.37 / 4.03 / 4.00 | 5.78 | 5.87 | 5.73 | 38.66 |
| Base | Profile | 4.63 / 6.00 / 5.21 | 4.21 / 3.97 / 4.00 | 5.91 | 5.96 | 5.76 | 40.08 |
| Base | CPG+Profile | 4.71 / 6.00 / 5.29 | 4.21 / 4.03 / 4.00 | 5.89 | 5.96 | 5.76 | 39.98 |
| GC | CPG | 4.53 / 6.00 / 5.26 | 4.47 / 4.00 / 4.00 | 5.89 | 6.01 | 5.77 | 40.46 |
| GC | Profile | 4.55 / 6.00 / 5.47 | 4.34 / 4.00 / 4.03 | 5.91 | 5.99 | 5.78 | 40.52 |
| GC (Selected) | CPG+Profile | 4.34 / 5.97 / 5.13 | 4.34 / 4.00 / 4.00 | 5.85 | 5.96 | 5.76 | 39.76 |
| GC + CoT | CPG+Profile | 4.42 / 6.00 / 5.39 | 4.32 / 4.00 / 4.00 | 5.96 | 6.19 | 5.78 | 30.72 |
| GC + MA (3 iter.) | CPG+Profile | 4.53 / 6.00 / 5.00 | 4.21 / 4.00 / 4.00 | 5.75 | 5.85 | 5.76 | 43.26 |
On the author-introduced multi-turn Therapy-Eval benchmark (100 Eeyore character cards, evaluated across up to 40 turns), Llama3-G2C exhibited sustained dialogic endurance:
| Model | Specificity (0โ4) (โ) | Competence (0โ4) (โ) | Authenticity (0โ4) (โ) | Safety (0โ4) (โ) | Flow (0โ4) (โ) | Average (โ) | Avg. Turns |
|---|---|---|---|---|---|---|---|
| CAMEL | 3.47*** | 3.22*** | 3.24*** | 4.00 | 3.95* | 3.58*** | 29.98 |
| Llama3-MAG | 3.27*** | 3.10*** | 2.69*** | 3.86** | 3.39*** | 3.26*** | 35.61 |
| Llama3-SQP | 3.73 | 3.48 | 3.77 | 4.00 | 4.00 | 3.80 | 22.56 |
| Llama3-G2C | 3.81 | 3.52 | 3.72 | 4.00 | 4.00 | 3.81 | 30.57 |
Key Findings¶
- Relational Grounding Elevates Realism and Flow: Expert clinicians favored Graph2Counsel for its natural conversational progression, noting that high-ranking dialogues delve deeply into 1โ2 tailored interventions (e.g., psychoeducation and homework planning) rather than superficially jumping between topics.
- Sustained Multi-Turn Therapeutic Alignment: In Therapy-Eval, Llama3-G2C sustained 30.57 turns on average while leading in competence (3.52) and specificity (3.81), whereas questionnaire-grounded baselines (SQPsychConv) decayed after 22.56 turns.
- Inherent Trade-Offs in CoT and Multi-Agent Prompts: Incorporating explicit Chain-of-Thought compressed overall conversation length from ~40 turns to ~30 turns due to verbosity, contradicting clinical preferences for unhurried sessions. Multi-Agent feedback doubled API costs with negligible quality increments.
Highlights & Insights¶
- Privacy-Preserving Structural Transfer: By converting identifiable clinical transcripts into topological symptom networks (CPGs) and re-synthesizing unconstrained client profiles, the framework achieves zero identity leakage while retaining 100% of the functional pathology structure.
- Broad Process-Based Therapy Coverage: Unlike existing datasets narrowly focused on standard CBT, Graph2Counsel operationalizes 257 distinct counselor strategies spanning ACT, DBT, motivational interviewing, and schema therapy.
- Pragmatic Prompt Optimization for Long-Form Generation: The finding that Guided Counseling matches or exceeds complex Multi-Agent and CoT workflows while preserving natural session length offers an efficient recipe for large-scale dialogue synthesis.
Limitations & Future Work¶
- Restricted Patient Diversity in Source Graphs: The 76 CPGs were derived from 6 patients treated at a single clinic for anxiety and depression, potentially encoding demographic and diagnostic biases into the underlying graphs.
- Static Graph Snapshots vs. Longitudinal Trajectories: Current CPGs represent single-session static states, unable to capture multi-week symptom evolution or therapeutic breakthroughs across longitudinal treatment.
- Evaluation Agent Clinical Alignment: Therapy-Eval utilizes off-the-shelf Llama3-8B and GPT-4o as client and judge agents, which may not capture nuanced psychiatric feedback as accurately as licensed practitioners.
Related Work & Insights¶
- vs CACTUS (Lee et al., 2024): CACTUS relies on crowdsourced persona descriptions from online forums; Graph2Counsel grounds generation in real clinical CPGs and expert-curated strategies, yielding superior authenticity (1.43 vs. 2.22).
- vs MDD-5k (Yin et al., 2025): MDD-5k samples external context triplets (time, person, event) from trees, whereas Graph2Counsel directly models the internal psychological mechanisms and causal interactions of the client.
- vs SQPsychConv (Vu et al., 2025): SQPsychConv conditions generation on coarse questionnaire ratings; Graph2Counsel leverages fine-grained CPG dependencies, allowing models to sustain longer, more competent sessions (30.57 vs. 22.56 turns).
Rating¶
- Novelty: โญโญโญโญโญ Pioneering integration of clinical psychological graphs (CPGs) and process-based multi-modality strategies for dialogue synthesis.
- Experimental Thoroughness: โญโญโญโญโญ Comprehensive blinded human evaluation by 4 licensed psychologists, 15 ablation configurations, 2 standard benchmarks, and an original multi-turn evaluation framework.
- Writing Quality: โญโญโญโญโญ Rigorous conceptual framing, clear methodological layout, and transparent analysis of prompt trade-offs.
- Value: โญโญโญโญโญ Provides a foundational methodology and open dataset for training safe, empathetic, and clinically competent mental health LLMs without privacy risks.