PaperlayerView exact revision on arXiv

Source-linked AI summary

PsyDT: Using LLMs to Construct the Digital Twin of Psychological Counselor with Personalized Counseling Style for Psychological Counseling

Haojie Xie, Yirong Chen, Xiaofen Xing, Jingkai Lin, Xiangmin Xu

arXiv:2412.13660v1cs.CL

TL;DR

Mental-health LLMs often overlook counselors’ distinct styles, and collecting real-world counseling cases is rare, costly, and time-consuming. PsyDT uses GPT-4 to synthesize style-specific multi-turn dialogues and fine-tune a counselor twin; the authors report approximately 70% similarity to real cases and better performance than other baselines.

  • Problem

    Mental-health LLMs overlook counselors’ differing linguistic styles and therapy techniques, although clients may seek different counseling styles; real-world multi-turn dialogue datasets are rare, and collecting counselor-specific cases is costly.

    Source: §1 · pp. 1–2

  • Method

    PsyDT uses GPT-4 to summarize counselor style and therapy techniques, synthesize multi-turn dialogues from single-turn cases, and fine-tune an LLM on PsyDTCorpus.

    Source: §1 · p. 2, §2 · p. 3

  • Results

    Similarity between PsyDT_Prompt dialogues and real-world counseling cases is approximately 70%, surpassing Smile_Prompt and SoulChat_Prompt, especially in therapy technique; PsyDT also outperforms other baselines.

    Source: §1 · p. 3, §3.1.1 · p. 5

  • Takeaways & Limitations

    PsyDT is presented as a quick, cost-effective way to produce a high-quality, style-specific multi-turn counseling corpus and counselor twin with potential for real-world application.

    Source: §5 · p. 8

    Without human feedback during fine-tuning, some model responses may pose potential harm; virtual conversational agents cannot replace real-world therapy.

    Source: Ethical Statement · p. 9

Abstract

from arXiv · show

Currently, large language models (LLMs) have made significant progress in the field of psychological counseling. However, existing mental health LLMs overlook a critical issue where they do not consider the fact that different psychological counselors exhibit different personal styles, including linguistic style and therapy techniques, etc. As a result, these LLMs fail to satisfy the individual needs of clients who seek different counseling styles. To help bridge this gap, we propose PsyDT, a novel framework using LLMs to construct the Digital Twin of Psychological counselor with personalized counseling style. Compared to the time-consuming and costly approach of collecting a large number of real-world counseling cases to create a specific counselor's digital twin, our framework offers a faster and more cost-effective solution. To construct PsyDT, we utilize dynamic one-shot learning by using GPT-4 to capture counselor's unique counseling style, mainly focusing on linguistic style and therapy techniques. Subsequently, using existing single-turn long-text dialogues with client's questions, GPT-4 is guided to synthesize multi-turn dialogues of specific counselor. Finally, we fine-tune the LLMs on the synthetic dataset, PsyDTCorpus, to achieve the digital twin of psychological counselor with personalized counseling style. Experimental results indicate that our proposed PsyDT framework can synthesize multi-turn dialogues that closely resemble real-world counseling cases and demonstrate better performance compared to other baselines, thereby show that our framework can effectively construct the digital twin of psychological counselor with a specific counseling style.

1 Introduction

Mental-health LLMs overlook counselors’ distinct linguistic and therapeutic styles, while collecting a specific counselor’s cases is costly and time-consuming. PsyDT uses GPT-4 to synthesize style-specific multi-turn dialogues and train a counselor twin; the authors report close resemblance to real cases and better performance than baselines.

1 Introduction

  • Existing mental-health LLMs overlook counselors’ differing linguistic styles and therapy techniques, although clients may seek different counseling styles.The introduction also says that fine-tuning on dialogues blending multiple styles often leads to response instability.Source: §1 · p. 2
  • Collecting many real-world counseling cases from one specific counselor is time-consuming and costly.Source: §1 · p. 2
  • PsyDT uses GPT-4 to simulate client personalities, summarize a counselor’s style and therapy techniques from real cases, synthesize multi-turn dialogues, and fine-tune on PsyDTCorpus.The framework selects 5,000 single-turn dialogues from SoulChatCorpus as synthesis inputs.Source: §1 · p. 2
  • PsyDT is presented as the first LLM-based framework for constructing a psychological counselor’s digital twin with personalized counseling style.Its synthesis method is described as a quick, cost-effective way to produce PsyDTCorpus, a style-specific multi-turn mental-health dialogue dataset.Source: §1 · p. 3
  • The authors report automatic and manual evaluations of both the synthetic dataset and the fine-tuned counselor model.Source: §1 · p. 3
  • PsyDT synthesizes multi-turn dialogues that closely resemble real-world counseling cases and performs better than other baselines.The authors frame these results as showing potential for real-world psychological counseling.Source: §1 · p. 3

2 Methodology

PsyDT synthesizes multi-turn dialogues through style and therapy-technique summarization, client-personality simulation, and dialogue generation, then fine-tunes a counselor-specific LLM on PsyDTCorpus.

2.1 Single-Turn Dialogues Preparation

  • GPT-4 is used to select 5,000 long-text single-turn dialogues from SoulChatCorpus; each includes a counseling topic, client-question title and details, and a counselor response.Source: §2.1 · p. 3

2.2.1 Real-World Counseling Cases Collection

  • The study collected real-world counseling cases with a professional counselor across 12 topics, with 12 cases in total and a reported total cost of $2000.Sessions were text-based and lasted 50 minutes; records were reviewed by the counselor and cleaned with GPT-4.Source: §2.2.1 · pp. 3–4

2.2.2 Linguistic Style and Therapy Technique Summarization

  • GPT-4 summarizes the counselor’s linguistic style and therapy technique from collected cases; all 12 cases used rational emotive behavior therapy (REBT).Source: §2.2.2 · p. 4

2.3 Client Personality Simulation

  • GPT-4 simulates clients’ Big Five personality traits from their questions to add complexity and diversity to clients’ linguistic styles.The five dimensions are Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism.Source: §2.3 · p. 4

2.4 Multi-Turn Mental Health Dialogues Synthesis

  • For multi-turn synthesis, single-turn dialogues are dynamically matched to real cases by counseling topic, combining topic-specific style, therapy technique, client traits, emotion guidance, and response guidance.GPT-4 uses this guidance to generate PsyDTCorpus.Source: §2.4 · p. 4

2.5 PsyDTLLM Model

  • PsyDTLLM is trained on PsyDTCorpus using multi-turn instruction fine-tuning, comparing each predicted counselor utterance with its target given the dialogue history.The training description defines the target and predicted utterances and the history of fewer than i turns.Source: §2.5 · p. 4

3 Experiments

The experiments analyze PsyDTCorpus and the fine-tuned PsyDTLLM through four research questions on similarity, dataset comparisons, synthesis components, and model performance.

  • How similar are synthetic multi-turn dialogues to real-world cases from a specific counselor?Source: §3 · p. 4
  • How does PsyDTCorpus perform compared with baseline datasets?Source: §3 · p. 4
  • Do the synthesis method’s components produce their corresponding effects in PsyDTCorpus?Source: §3 · p. 4
  • How does PsyDTLLM perform compared with baseline language models?Source: §3 · p. 4

3.1 Analysis of PsyDTCorpus Dataset

The analysis compares PsyDTCorpus with real-world counseling cases and baseline datasets, then tests whether its synthesis method integrates counselor style and client personality.

3.1.1 RQ1: Similarity Analysis

  • PsyDT_Prompt achieves approximately 70% similarity to real-world counseling cases, surpassing Smile_Prompt and SoulChat_Prompt, especially in therapy technique.Source: Figure 4 · p. 5, §3.1.1 · p. 5

3.1.2 RQ2: Overall Dataset Comparisons

  • For the dataset comparison, five evaluators rated 50 samples from each dataset, with moderate agreement across the reported assessment dimensions.The evaluators were four senior psychological postgraduate students and an experienced psychotherapist; Fleiss’ κ values ranged from 0.403 to 0.547.Source: §3.1.2 · p. 5
  • PsyDTCorpus significantly outperforms other datasets across all four professional assessment dimensions.The dimensions are Conversation Strategy, State and Attitude, Relationship Building, and Application of Therapy Technique; 50 dialogues per dataset were manually evaluated.Source: §3.1.2 · pp. 5–6, Figure 5 · p. 6

3.1.3 RQ3: Ablation Studies

  • The synthesized dialogues achieve over 60% fidelity when evaluated against ablated versions, supporting integration of linguistic style, therapy technique, and client personality.The ablations each excluded one of those three elements.Source: Figure 6 · p. 6, §3.1.3 · p. 6

3.2 Analysis of PsyDTLLM model

PsyDTLLM is compared with closed-source, open-source, and domain-specific baselines; it outperforms the baselines on reported automatic and professional evaluations.

3.2.1 Baselines

  • Baselines include closed-source, open-source, and domain-specific models, with open-source comparisons selected at similar 6B–9B parameter sizes.Source: §3.2.1 · pp. 6–7

3.2.2 Implementation details

  • Qwen2-7B-Instruct is the backbone, and PsyDTLLM is fine-tuned on PsyDTCorpus for three epochs.Training uses batch size 2 per GPU on eight NVIDIA-A800-80G GPUs.Source: §3.2.2 · p. 7

3.2.3 RQ4: Overall Model Comparisons

  • PsyDTLLM outperforms the baseline models in both automatic evaluation metrics and professional metrics.Automatic evaluation uses ROUGE-1, ROUGE-2, ROUGE-L, BLEU-4, and FBERT; professional evaluation covers empathy, conversation strategy, state and attitude, and safety.Source: §3.2.3 · p. 7

3.3 Case Study

The case study contrasts how five models respond to clients seeking emotional support and specific counseling styles.

  • PsyDTLLM implicitly expresses emotional and cognitive empathy, while ChatGPT and Qwen2-7B-Instruct tend to advise rather than ask questions or listen.SoulChat does not meet clients’ specific style needs, while CPsyCounX’s explicit therapy terminology may confuse clients outside psychology.Source: §3.3 · p. 7

4 Related Work

Related work covers multi-turn dialogue generation, empathetic counseling datasets, and methods for reconstructing counseling conversations. PsyDT focuses on constructing a counselor twin with a personalized style, while the therapy-technique discussion describes REBT, CBT, and DBT.

4.1 Mental Health LLMs

  • SMILE expands single-turn dialogues into multi-turn interactions, while SoulChatCorpus contains over 2 million samples for empathetic dialogue fine-tuning.SMILE is attributed to Qiu et al. (2024a), and SoulChatCorpus to Chen et al. (2023).Source: §4.1 · p. 7
  • CPsyCoun reconstructs and evaluates Chinese counseling dialogues using two-phase generation and a multi-turn counseling benchmark; PsyDT targets personalized counselor style.CPsyCoun is attributed to Zhang et al. (2024).Source: §4.1 · p. 8

4.2 Therapy Techniques

  • REBT links emotional distress and dysfunctional behavior to irrational beliefs and people’s interpretations and reactions to events.REBT is attributed to Dryden (2005).Source: §4.2 · p. 8
  • CBT, developed primarily by Aaron Beck in the 1960s, focuses on interplay among thoughts, emotions, and behaviors; DBT was initially designed for borderline personality disorder and later adapted to other conditions.Source: §4.2 · p. 8

5 Conclusion

PsyDT proposes a framework for constructing a psychological counselor’s digital twin with a personalized counseling style. Its synthesis method produces a high-quality multi-turn dialogue dataset that closely resembles real-world counseling cases, indicating potential for real-world counseling applications.

5 Conclusion

  • PsyDT proposes an LLM framework for constructing a psychological counselor’s digital twin with a personalized counseling style.The framework’s dialogue synthesis method can quickly and cost-effectively produce PsyDTCorpus.Source: §5 · p. 8
  • PsyDTCorpus is a high-quality multi-turn mental health dialogue dataset whose dialogues closely resemble real-world counseling cases.The conclusion describes its synthesis as quick and cost-effective.Source: §5 · p. 8
  • The authors state that PsyDT’s results indicate potential for application in real-world psychological counseling.Source: §5 · p. 8

Limitation and Future Work

PsyDT constructs a digital twin with a specific counseling style, but this does not guarantee that it will solve clients’ psychological problems or meet all counseling needs. The authors identify reliance on one counselor as a limitation and propose exploring multiple counselors in joint diagnosis.

Limitation and Future Work

Ethical Statement

The authors describe data-cleaning steps intended to protect privacy and remove potentially harmful conversations. They also note that fine-tuning did not include human feedback and caution that the model is not a substitute for professional therapy.

Ethical Statement

  • Rule-based cleaning, manual rewriting, and human proofreading were used to remove sensitive personal information from the dataset.The authors also report expunging conversations with potential harm to clients, others, or society.Source: Ethical Statement · p. 9
  • Without human feedback during fine-tuning, some PsyDTLLM responses may pose potential harm to users.The authors recommend seeking a professional counselor or psychiatrist if user interactions do not noticeably improve.Source: Ethical Statement · p. 9
  • Virtual conversational agents cannot replace real-world therapy; downstream AI responses should be used solely as references.Source: Ethical Statement · p. 9
  • Psychology annotators received $0.418 per sample, equivalent to an hourly wage of $8.36, above the stated U.S. federal minimum wage of $7.25.Each annotator spent approximately three minutes assessing each sample.Source: Ethical Statement · p. 9

A Reproducibility Checklist

The checklist specifies PsyDTCorpus/PsyDTLLM release timing, training hardware, driver versions, and recommended software packages.

A Reproducibility Checklist

B Results of Model Automatic Evaluation

This section covers automatic-evaluation results using ROUGE and BLEU across 12 counseling topics, along with evaluation prompts, example counseling exchanges, and an algorithm for synthesizing multi-turn dialogues.

B Results of Model Automatic Evaluation

  • Figure 7 covers ROUGE and BLEU results for PsyDTLLM and baseline models across 12 counseling topics, but no scores or comparative winner are provided here.The text says Figure 7 contains four semantic metrics, while its caption specifies ROUGE and BLEU.Source: B Results of Model Automatic Evaluation · p. 11, Figure 7 · p. 12
  • Figure 8 is identified as an example of the single-turn long-text dialogues used in the paper.The passages show a client describing low class-activity attendance and a counselor asking clarifying questions.Source: Figure 8 · p. 13, B Results of Model Automatic Evaluation · p. 14
  • Figure 16 identifies an example of PsyDTCorpus, distinguishing client-personality, counselor-linguistic-style, and therapy-technique segments.The supplied example includes a counselor reflecting the client's feelings and asking follow-up questions.Source: B Results of Model Automatic Evaluation · p. 22, Figure 16 · p. 23
  • The evaluation prompts score responses against an evaluation standard, conversation history, reference answer, and poor answer, with an expert evaluator role and prompts for strategy, state and attitude, and safety.The evaluator is specified as an expert in psychology and linguistics.Source: B Results of Model Automatic Evaluation · pp. 28–29, Figure 21 · p. 28, Figure 22 · p. 29
  • PsyDT’s synthesis algorithm uses GPT-4 to turn single-turn dialogues into multi-turn dialogues conditioned on counselor style and therapy technique, topic cases, emotion and response guides, and simulated client Big Five traits.Its inputs include a list of 5,000 single-turn dialogues and counselor cases covering 12 topics.Source: B Results of Model Automatic Evaluation · p. 33
  • The examples include counseling exchanges about anxiety and academic pressure, including a PsyDTLLM case in which a client describes graduate-school distress and coping by isolating.The displayed dialogue includes counselor responses acknowledging pressure and asking about the client’s experience and coping.Source: B Results of Model Automatic Evaluation · pp. 35–36, Figure 28 · p. 36

Related papers

Loading 2412.13660v1…