DIALOGUE GENERATIVE MODEL FOR MENTAL THERAPY

Abstract

The increasing demand for accessible mental health re- sources has inspired the development of intelligent systems capable of offering immediate support. In this study, we present a mental health chatbot designed to provide empathetic and contextually relevant assistance to users. We initiated our approach by employing foundational NLP techniques, related to open-domain question answering methods, to establish a baseline understanding of conversational dynamics. Subse- quently, we explored the application of generative models to capture the nuanced and sensitive nature of mental health queries. Our methodology started with the full fine-tuning of smaller models, which provided initial insights into the training process of LLMs. Building upon this foundation, we leveraged the Low-Rank Adaptation (LoRA) [1] technique to fine-tune larger, more complex models, thus harnessing their superior generative capabilities without the computational expense of training from scratch. The success of our models can be attributed to the properly formatted and curated dataset, which was crucial in training the chatbot to understand and respond to a diverse range of mental health queries. The final chatbot demonstrates promising capabilities in delivering instant, reliable, and empathetic interaction, marking a signifi- cant step forward in digital mental health assistance. Our work not only showcases the potential of hybrid NLP applications in mental health scenarios but also paves the way for further innovations in therapeutic conversational agents.