Roleplaying with Structure: Synthetic Therapist-Client Conversation Generation from Questionnaires

Doan Nam Long Vu, Rui Tan, Lena Moench, Svenja Jule Francke, Daniel Woiwod, Florian Thomas-Odenthal, Sanna Stroth, Tilo Kircher,, Christiane Hermann, Udo Dannlowski, Hamidreza Jamalabadi, Shaoxiong Ji, Technical University of Darmstadt, Philipps-University Marburg, Justus Liebig University Giessen, · 2025

Read Roleplaying with Structure: Synthetic Therapist-Client Conversation Generation from Questionnaires by Doan Nam Long Vu, Rui Tan, Lena Moench, Svenja Jule Francke, Daniel Woiwod, Florian Thomas-Odenthal, Sanna Stroth, Tilo Kircher,, Christiane Hermann, Udo Dannlowski, Hamidreza Jamalabadi, Shaoxiong Ji, Technical University of Darmstadt, Philipps-University Marburg, Justus Liebig University Giessen, (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2510-25384.txt.

Introduction

Doan Nam Long Vu led twelve co-authors on synthetic therapy data. Rui Tan, Lena Moench, and Svenja Jule Francke helped. Daniel Woiwod, Florian Thomas-Odenthal, and Sanna Stroth followed. Tilo Kircher, Christiane Hermann, and Udo Dannlowski joined. Hamidreza Jamalabadi, Simone Balloccu, and Shaoxiong Ji followed. Their 2025 paper generates therapist-client conversations with language models.

Privacy rules limited earlier mental health data work. Prior studies relied mainly on generic information instead. Their SQPsych pipeline uses structured questionnaire based profiles. Real client profiles inform generation without leaking sensitive data. The resulting corpus is called SQPsychConv in the paper. They fine-tune open-weight language models on it. Automatic benchmarks and human evaluation tested the models. Trained psychotherapists judged quality in human evaluation. Standard benchmarks missed strengths of their data set. Expert judgment showed clear gains in therapist roleplaying. Experts preferred their sessions over rival mental health models.

Catalog record

CREATORDoan Nam Long Vu, Rui Tan, Lena Moench, Svenja Jule Francke, Daniel Woiwod, Florian Thomas-Odenthal, Sanna Stroth, Tilo Kircher,, Christiane Hermann, Udo Dannlowski, Hamidreza Jamalabadi, Shaoxiong Ji, Technical University of Darmstadt, Philipps-University Marburg, Justus Liebig University Giessen,
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE67 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2510-25384

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Roleplaying with Structure: Synthetic Therapist-Client Conversation Generation from Questionnaires Doan Nam Long Vu, Rui Tan, Lena Moench, Svenja Jule Francke, Daniel Woiwod, Florian Thomas-Odenthal, Sanna Stroth, Tilo Kircher,, Christiane Hermann, Udo Dannlowski, Hamidreza Jamalabadi, Shaoxiong Ji, Technical University of Darmstadt, Philipps-University Marburg, Justus Liebig University Giessen, University of Münster, ELLIS Institute Finland, University of Turku https://ai-mh.github.io/ Abstract The development of AI for mental health is hindered by a lack of authentic therapy dialogues, due to strict privacy regulations and the fact that clinical sessions were historically rarely recorded. We present an LLM-driven pipeline that generates synthetic counseling dialogues based on structured client profiles and psychological questionnaires. Grounded on the principles of Cognitive Behavioral Therapy (CBT), our method creates synthetic therapeutic conversations for clinical disorders such as anxiety and depression. Our framework, SQPsych (Structured Questionnaire-based Psychotherapy), converts structured psychological input into natural language dialogues through therapist-client simulations. Due to data governance policies and privacy restrictions prohibiting the transmission of clinical questionnaire data to third-party services, previous methodologies relying on proprietary models are infeasible in our setting. We address this limitation by generating a high-quality corpus using open-weight LLMs, validated through human expert evaluation and LLM-based assessments. Our SQPsychLLM \textrm{SQPsychLLM}_{\text{}} models fine-tuned on SQPsychConv \textrm{SQPsychConv}_{\text{}}^{\texttt{}} achieve strong performance on counseling benchmarks, surpassing baselines in key therapeutic skills. Our findings highlight the potential of synthetic data to enable scalable, data-secure, and clinically informed AI for mental health support. We will release our code, models, and corpus at https://ai-mh.github.io/SQPsych 1 Introduction Figure 1: Our pipeline begins with client information and questionnaire data in numerical form. We then convert this into natural language, which is used to simulate therapist-client roleplay conversations. Acquiring therapy-client dialogues from clinical institutions is difficult due to strict privacy laws such as HIPAA (US), GDPR (EU) and PIPL (China), along with ethical issues regarding patient confidentiality De Freitas et al. 2022 . These restrict access to large and varied datasets necessary to build AI systems for mental health care. Furthermore, clinical psychology interviews and counseling sessions from the past were not routinely recorded or transcribed. The questionnaires in practice were often derived from informally distilled insights during these sessions. With LLM, there is now potential to reuse these processes and data sources to synthetically generate more structured and meaningful outcomes, such as therapy-client dialogues. Synthetic therapy-client conversations offer valuable opportunities in mental health care and clinical research Sharma et al. 2021 . They can also power AI tools for screening, and diagnostic support, helping to streamline the intake process and reduce pressure on already overburdened therapists Cabrera et al. 2023 . In addition, simulated therapy sessions can serve as a training tool for students and early career mental health and social workers, providing hands-on experience in a controlled, ethical environment Fu et al. 2023 . We introduce an LLM-based generation pipeline that produces synthetic counseling dialogues conditioned on structured client data and standardized mental health assessments, and psychologically informed guidance. These conversations are tailored to mental health conditions (e.g., major depressive disorder and anxiety). By integrating expertise and best practice in psychology (Cognitive Behavior Therapy - CBT) Beck 1963 , the resulting data can serve as a valuable resource to improve fine-tuning therapeutic agents, intervention planning. By bridging structured psychological data and natural language conversation, our approach contributes to scalable and clinically relevant conversation generation. Figure 1 demonstrates our main concept. Our contributions span modeling, data, and evaluation: • Modeling - SQPsych : We introduce SQPsych ( S tructured Q uestionnaire-based Psych otherapy), a novel pipeline that transforms structured client metadata and questionnaires into free text, employs role-playing between therapist and client personas, enabling generation of counseling conversations. • Data & Model Resources - Synthetic Corpus & Fine-tuned Models : We use seven open models (23B–123B) for local synthetic data generation, conditioned on real-world structured client information and questionnaires. This choice ensures sensitive questionnaires are not shared with proprietary models. We then leverage the synthetically generated dataset, distilled from larger LLMs, to fine-tune seven smaller models, enabling resource efficiency and practical utility. • Evaluation & Insight - Benchmarking & Expert Feedback : We perform a holistic evaluation with human experts and automatic metrics. (1) Our models perform effectively on two specialized counseling skill benchmarks, exceeding baseline models in key therapeutic competencies such as distortion and core beliefs’ identification. (2) In preference evaluations conducted by both an LLMs panel and human experts, our fine-tuned model consistently outperforms the baselines. (3) Expert feedback offers valuable insights for enhancing future data/models focused on counseling tasks. 2 Methods We propose SQPsych , a pipeline converting structured data to conversations which synthesizes clinically therapist-client dialogues by conditioning on structured mental health data. This pipeline integrates standardized client metadata and questionnaire, combined with targeted role-based generation strategies informed by clinical expertise. We describe the most important aspect of our pipeline in this section. Detailed information on the generation workflow can be found in the Appendix B . 2.1 Structured Data Conditioning & Conversion Our pipeline is conditioned on structured information: general client information (e.g., age, gender, etc.) and standardized questionnaire results (e.g., PHQ-9, BDI, etc.). Since much of this data is available in numerical or categorical form, we convert it to natural language for proper descriptions. This conversion may help the model interpret client characteristics and psychological states with more precision. We construct a hash map that maps numerical values to their corresponding natural language descriptions for this purpose. Figure 2 shows an example of the formatting of our data input. Figure 2: An artificial example of converting numerical values, i.e., structured client metadata (e.g., Age, Age of Onset, MDD_Current) and questionnaire scores (e.g., BDI1, BDI2 and HAMD1, HAMD2), into natural language with structure to natural language conversion. 2.2 Dual-role Generation Strategy Previous approaches, such as CACTUS Lee et al. 2024 and SMILE ( Qiu et al. 2024 ) , generate conversations using a single model that simulates both roles simultaneously. In contrast, we employ a dual-agent framework where separate models independently simulate the therapist and client, each operating with distinct reasoning processes and information access patterns. This approach enables dialogue dynamics, as each agent maintains its own perspective and responds based on role-specific knowledge and objectives. Specifically, the client agent is conditioned on background information and questionnaire data to generate responses that reflect the client’s psychological state and personal circumstances. The therapist agent operates with therapeutic guidelines and intervention strategies, responding to the client’s statements. For that in the generated dialogue, the LLM therapist will apply CBT skills (e.g., paraphrasing, validation ), while the LLM client will express emotions and hesitation , grounded in personal data and questionnaire-derived. 2.3 Turn-Based Generation Workflow The conversation proceeds in a role-playing therapeutic simulation.The workflow includes the following steps: 1. Initialization. We initialize two language models, one for the role of the therapist and the other for the client. 2. Turn-by-turn interaction. The interaction proceeds turn-by-turn: the therapist model generates an initial utterance, which is then passed to the client model to produce a response. This exchange continues iteratively, and both models incorporate the full conversation history at each step to maintain contextual consistency. 3. Termination criterion. The dialogue progresses until the therapist model issues an explicit signal with an

Read the complete text