Note2Narr: a neuro-symbolic multi-agent system for synthesizing high-fidelity training narratives from sparse counselor notes

Abstract

Introduction Early identification of mental health risks among university students requires high-quality, privacy-conscious dialogue data for developing and evaluating computational screening models. Existing synthetic datasets often lack the contextual nuance and diagnostic consistency needed to preserve clinically relevant psychological signals. Methods We introduce Note2Narr, a neuro-symbolic framework for synthesizing high-fidelity training narratives from sparse counselor notes. A Dynamic Counseling Tree guides inquiry from broad topic exploration to in-depth probing of personal experiences. A Supervisor Agent equipped with a Chain of Psychological Clues audits the safety and consistency of generated narratives, while a Fictional Detail Generator incorporates diverse, disorder-neutral personal experiences. A Vendi Score-guided 1-to-X strategy adaptively allocates the number of narrative variants. Evaluation combines blinded human assessment, LLM-based assessment, student-disjoint downstream prediction, and a label-leakage audit with masking ablation. Results The framework produced 981 multi-turn counseling dialogues comprising 11,145 question-answer turns from 224 source students. Blinded human comparisons with MDD-5K identified statistically significant improvements in Professionalism, Communication, Participation, Repetition, and Authenticity after Holm correction. Downstream text classifiers achieved macro-AUROC and macro-AUPRC values of up to 0.739 and 0.405, respectively. Masking explicit diagnostic cues changed macro-AUPRC by at most 0.005, suggesting that predictive utility was not driven solely by direct diagnostic wording. Discussion These findings support Note2Narr as a framework for constructing synthetic counseling corpora for privacy-conscious dialogue modeling and screening-oriented research. The results demonstrate learnable associations in synthetic narratives; prospective and external validation remains necessary before applied use.