Conversation Analysis of Remote Dialogue System for Mental Health Interventions

Mental health conditions significantly impact all aspects of life, highlighting the need for innovative and accessible interventions. Remote automatic dialogue systems are increasingly used in mental health interventions to help individuals express their concerns and relieve stress. We used NLP technologies to address three key challenges in mental health intervention dialogue systems: identifying user concern topics, predicting concern severity, and optimizing dialogue flow. First, we introduced a new topic clustering method to classify samples into customized categories without the need for large labeled datasets, achieving an accuracy of 73.8%. Next, we created a dataset containing user concerns and severity labels to train models for detecting severe concerns. In our evaluation, BERT achieved a Pearson correlation coefficient of 0.83 in predicting concern severity. We also used language models for dialogue flow optimization. BERT achieved an F-score of 0.752 in predicting user satisfaction. We continuously monitored and predicted user satisfaction based on conversation content, identifying flows that negatively impacted satisfaction. We believe that our AI-supported dialogue system, capable of continuous adaptation and improvement, can significantly contribute to mental health interventions.