Multi-View Attention Multiple-Instance Learning Enhanced by LLM Reasoning for Cognitive Distortion Detection
Cognitive distortions, often associated with various mental health disorders, pose significant challenges in their automatic detection due to factors such as contextual ambiguity, co-occurrence, and semantic overlap. In a groundbreaking study published as arXiv:2509.17292v3, researchers have introduced a novel framework that integrates Large Language Models (LLMs) with a Multiple-Instance Learning (MIL) architecture. This innovative approach aims to enhance interpretability and expression-level reasoning in identifying cognitive distortions.
Framework Overview
The proposed framework decomposes each utterance into three key components: Emotion, Logic, and Behavior (ELB). These components are then processed by LLMs to infer multiple distortion instances. Each instance is categorized with a predicted type, its expression, and a model-assigned salience score. The integration of these instances is facilitated through a Multi-View Gated Attention mechanism, which ultimately aids in the final classification of cognitive distortions.
Key Features of the Framework
- Emotion, Logic, and Behavior Components: The decomposition into ELB allows for a more nuanced understanding of the utterances, making it easier to identify and categorize cognitive distortions.
- Large Language Models: Leveraging LLMs enhances the processing capabilities, allowing for improved inference of distortion instances.
- Multi-View Gated Attention Mechanism: This mechanism integrates the various distortion instances, facilitating a cohesive classification output.
Experimental Results
The researchers conducted experiments using two distinct datasets: the Korean dataset (KoACD) and the English dataset (Therapist QA). The findings indicate that incorporating the ELB components alongside LLM-inferred salience scores significantly boosts classification performance. This improvement is particularly notable for cognitive distortions characterized by high interpretive ambiguity, underscoring the effectiveness of the proposed framework.
Implications for Mental Health NLP
The results of this study suggest a psychologically grounded and generalizable approach to fine-grained reasoning within the realm of mental health natural language processing (NLP). By addressing the complexities inherent in cognitive distortion detection, this framework holds promise for advancing automatic detection methodologies, which could ultimately enhance therapeutic interventions and support systems for individuals facing mental health challenges.
Access to Resources
The dataset and implementation details of this innovative framework are publicly accessible, encouraging further research and exploration in this critical field. The study not only contributes to the academic discourse on cognitive distortions but also paves the way for future advancements in mental health analysis through AI technologies.