Grandes modelos de linguagem aplicados a intervenções em saúde mental

Abstract

Growing concern about the need for effective mental health interventions is evident, especially given the shortage of qualified professionals to deal with the existing demand. The mental health field often finds itself in a situation where demand for services significantly outstrips supply, resulting in many individuals not having access to the necessary support. Faced with this scenario, it is crucial to explore viable alternatives to help provide care. Among the various demands present in the area of mental health in Brazil, alcohol abuse emerges as one of the major problems faced as a public health issue. A possible solution to this problem arises with the use of large language models to understand and generate text in a contextually relevant and accurate way. These models can play a significant role in expanding access to mental health services. This approach can not only help relieve pressure on mental health professionals, but also reach a larger portion of the population who may not have easy access to therapists or specialised clinics. This dissertation explores the proposal of an intervention aid tool for mental health, especially in the context of alcohol use, providing help to the counsellor to follow best practices. In order to better understand the tools and look for the most appropriate ways forward, a systematic review was carried out. This review aimed to map the characteristics of chatbots aimed at mental health, including aspects of usability, technologies and methodologies used, evaluation methods and the stage of development of studies. Based on these findings, the dissertation proposes to explore the application of the major language models LLAMA and GEMMA in the context of brief intervention on Alcohol and Health. Firstly, modelling is carried out using publicly available data. This validates our proposed solution by delivering satisfactory results, even when using data translated from English to train the model. Secondly, the process of training the model and adapting it to meet our objectives is demonstrated. We used data from real therapy sessions. The models are evaluated by expert evaluators in the problem domain. This work contributes to understanding the current state of chatbots applied to mental health. In addition, the comparison and refinement of the LLAMA and GEMMA models promotes understanding of the performance of language models in specific mental health contexts