Crisp: Cognitive Restructuring of Negative Thoughts through Multi-turn Supportive Dialogues (2504.17238v1)

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.

Cognitive Restructuring Through Supportive Dialogue in AI

This paper presents "Crisp," a framework leveraging Cognitive Restructuring (CR) techniques for transforming individuals' negative thoughts into positive outlooks via multi-turn dialogues with AI. The authors aim to bridge a critical gap in psychotherapeutic services, particularly given current clinician shortages and the stigmatization of mental health treatment, by utilizing LLMs in interactive psychotherapy.

The CRDial framework is central to this paper, facilitating authentic therapeutic dialogue through a multichannel, multi-stage process. CRDial consists of identification, restructuring, and support phases to guide dialogue. Unlike simpler text rewriting methods or fixed-pattern dialogue approaches in prior studies, CRDial aims to model therapy's iterative nature, which addresses diverse and persistent cognitive distortions in a comprehensive manner. This framework employs strategies at the sentence level, inspired by Hill's Help Skill Theory and Dialectical Behavior Therapy, ensuring that the dialogue not only restructures thoughts but also provides emotional support.

The authors constructed Crisp, a bilingual dataset featuring high-quality dialogues distilled from LLMs, specifically GPT-4o. The dataset comprises 22,000 dialogues covering a range of situations where individuals face mental health challenges. It provides strategy labels and cognitive distortion annotations, undergoing rigorous quality control to ensure the data's relevance, safety, and professional integrity. The formation of Crisp demonstrates effectiveness across varied cognitive styles and thought patterns, owing to its multi-channel loop mechanism which identifies and restructures multiple cognitive distortions within a single issue.

Crispers models were trained on Crisp, leveraging dual objectives for strategy-controlled generation and distortion identification. Crispers, available in 7B and 14B scales, were evaluated extensively and showed superior performance in various mental health situations, outperforming baseline models, including the teacher model GPT-4o. The evaluation involved interactive pointwise and pairwise analysis across diverse mental health situations like family and education, highlighting Crispers' robust potential in addressing individual cognitive styles through supportive dialogue.

The paper includes a psychological intervention trial that quantifies the affective shifts in participants after interaction with different models. Crispers-14B notably enhances positive affect scores and alleviates negative affect scores more effectively compared to GPT-4o and Emohaa. This suggests a promising avenue for AI in providing clinical psychotherapy support.

This research has significant implications, advancing human-LLM interactive psychotherapy for CR. The proposed solution not only enhances accessibility to mental health resources but also augments theoretical frameworks for modeling emotional intelligence and adaptive feedback within AI systems. Future developments may explore more scalable applications of Crispers across diverse domains or examine the integration of CRDial framework functionalities into existing therapeutic practices.

The paper elucidates a promising convergence between psychological counseling and AI, sparking further research on interactive dialogue systems' capabilities in cognitive behavioral therapy. As AI technology continues to progress, the integration of nuanced emotional understanding and adaptive strategies within conversational agents could redefine mental health support paradigms.