Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Crisp: Cognitive Restructuring of Negative Thoughts through Multi-turn Supportive Dialogues

Jinfeng Zhou, Yuxuan Chen, Jianing Yin, Yongkang Huang, Yihan Shi, Xikun Zhang, Libiao Peng, Rongsheng Zhang, Tangjie Lv, Zhipeng Hu, Hongning Wang, Minlie Huang

TL;DR

This work introduces CRDial, a comprehensive framework for cognitive restructuring implemented through multi-turn, emotion-supportive dialogues. It combines CT-guided identification of automatic, intermediate, and core thoughts with DAT-driven restructuring, underpinned by a multi-channel loop and external commonsense knowledge to address diverse cognitive distortions. The authors distill Crisp, a 22k bilingual CR dialogue dataset, and train Crispers, CR-focused LLMs at 7B and 14B, showing through human evaluations that Crispers outperform a teacher model on several metrics and that Crispers can meaningfully improve affect in a clinical-like intervention. The results suggest CRDial offers a scalable, high-quality approach to human-LLM interactive psychotherapy for CR, with strong implications for accessibility and cross-linguistic applicability, while acknowledging limitations and the need for broader clinical validation.

Abstract

Cognitive Restructuring (CR) is a psychotherapeutic process aimed at identifying and restructuring an individual's negative thoughts, arising from mental health challenges, into more helpful and positive ones via multi-turn dialogues. Clinician shortage and stigma urge the development of human-LLM interactive psychotherapy for CR. Yet, existing efforts implement CR via simple text rewriting, fixed-pattern dialogues, or a one-shot CR workflow, failing to align with the psychotherapeutic process for effective CR. To address this gap, we propose CRDial, a novel framework for CR, which creates multi-turn dialogues with specifically designed identification and restructuring stages of negative thoughts, integrates sentence-level supportive conversation strategies, and adopts a multi-channel loop mechanism to enable iterative CR. With CRDial, we distill Crisp, a large-scale and high-quality bilingual dialogue dataset, from LLM. We then train Crispers, Crisp-based conversational LLMs for CR, at 7B and 14B scales. Extensive human studies show the superiority of Crispers in pointwise, pairwise, and intervention evaluations.

Crisp: Cognitive Restructuring of Negative Thoughts through Multi-turn Supportive Dialogues

TL;DR

This work introduces CRDial, a comprehensive framework for cognitive restructuring implemented through multi-turn, emotion-supportive dialogues. It combines CT-guided identification of automatic, intermediate, and core thoughts with DAT-driven restructuring, underpinned by a multi-channel loop and external commonsense knowledge to address diverse cognitive distortions. The authors distill Crisp, a 22k bilingual CR dialogue dataset, and train Crispers, CR-focused LLMs at 7B and 14B, showing through human evaluations that Crispers outperform a teacher model on several metrics and that Crispers can meaningfully improve affect in a clinical-like intervention. The results suggest CRDial offers a scalable, high-quality approach to human-LLM interactive psychotherapy for CR, with strong implications for accessibility and cross-linguistic applicability, while acknowledging limitations and the need for broader clinical validation.

Abstract

Cognitive Restructuring (CR) is a psychotherapeutic process aimed at identifying and restructuring an individual's negative thoughts, arising from mental health challenges, into more helpful and positive ones via multi-turn dialogues. Clinician shortage and stigma urge the development of human-LLM interactive psychotherapy for CR. Yet, existing efforts implement CR via simple text rewriting, fixed-pattern dialogues, or a one-shot CR workflow, failing to align with the psychotherapeutic process for effective CR. To address this gap, we propose CRDial, a novel framework for CR, which creates multi-turn dialogues with specifically designed identification and restructuring stages of negative thoughts, integrates sentence-level supportive conversation strategies, and adopts a multi-channel loop mechanism to enable iterative CR. With CRDial, we distill Crisp, a large-scale and high-quality bilingual dialogue dataset, from LLM. We then train Crispers, Crisp-based conversational LLMs for CR, at 7B and 14B scales. Extensive human studies show the superiority of Crispers in pointwise, pairwise, and intervention evaluations.

Paper Structure

This paper contains 71 sections, 3 equations, 4 figures, 24 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. CRDial Framework
  4. Supportive Conversation Strategies
  5. Supportive Strategy Scheme
  6. Sentence-Level Strategy Constraints
  7. CT-guided Identification Stage
  8. Help-seeking Situation Collection
  9. CT-guided Identification Process
  10. Implementation
  11. DAT-driven Restructuring Stage
  12. DAT-driven Restructuring Process
  13. Implementation
  14. Multi-channel Loop Mechanism
  15. Multi-channel Mechanism
  16. ...and 56 more sections

Figures (4)

  • Figure 1: A data example of Crsip crafted via LLMs using CRDial, which identifies and restructures multiple negative thoughts (i.e., cognitive distortions) via multi-turn dialogue with emotional support.
  • Figure 2: CRDial framework used to distill dialogues for LLMs, clarified in the left.
  • Figure 3: Distributions of mental health situations in our Crisp, with 10 categories across 54 sub-categories.
  • Figure 4: Positive and negative affect changes of the psychological intervention trial. Error bars show bootstrapped 95% confidence intervals.