High-Quality Dataset Construction and Large Language Model Adaptation for Japanese Mental Health Response Generation

​

Large language models (LLMs) have shown considerable potential for supporting mental health-related communication. However, large-scale Japanese datasets suitable for response generation remain limited, and the influence of training data quality on generated responses has not been sufficiently investigated. This study constructs a dataset for Japanese mental health response generation using conversations collected from an anonymous online counseling forum. Personally identifiable information, AI-generated responses, self-replies, extremely short responses, and other unsuitable samples are removed to improve the reliability of the training data. Responses highly rated by users are then utilized to adapt an LLM to Japanese mental health-related conversations. The adapted model is compared with a general-purpose model to examine how data selection and quality-aware training affect response generation. The generated responses are quantitatively evaluated using multiple automatic metrics. This study aims to provide insights into dataset construction and model adaptation for generating more natural and useful responses in Japanese mental health support settings.