Entendendo sintomas de depressão em redes sociais: uma abordagem de granularidade fina com volume de dados restrito

• Ano de defesa: 2024
• Autor(a) principal: Mendes, Augusto Rozendo
• Orientador(a): Caseli, Helena de Medeiros
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de São Carlos; Câmpus São Carlos
• Programa de Pós-Graduação: Programa de Pós-Graduação em Ciência da Computação - PPGCC
• Palavras-chave em Português: Depressão; PLN; Aprendizado multi-tarefa; Aprendizado de máquina; Redes sociais
• Palavras-chave em Inglês: Depression; NLP; Multi-task learning; Machine learning; Low-resource; Social media
• Área do conhecimento CNPq: CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO::METODOLOGIA E TECNICAS DA COMPUTACAO
• Link de acesso: https://repositorio.ufscar.br/handle/20.500.14289/21150

This study investigated the identification of depression signs in online text, utilizing a set of fine-grained labels, composed of 21 distinct signs, in order to deep the collective understanding of how depression is expressed online. Results indicated that emotional and external signs of depression are frequent in social media, while somatic signs are scarcely expressed; however this trend does not carry over to model performance, with models performing best in somatic sign classification and struggling with some of the most frequent signs. Given these challenges regarding model performance, potentially related to data scar- city, a series of techniques were evaluated with the goal of improving model performance, including regularization techniques, data augmentation, prompt engineering and multi- task learning, among which multi-task learning proved to be the most promising. With the continuation of joint learning experiments, additional research questions concerning which auxiliary tasks lead to positive transfer - and why - were answered: 3 of the 7 auxiliary tasks led to positive transfer, including depression sign classification under a simplified taxonomy, fine-grained emotion classification and sentiment classification led to positive transfer, however none of a set of 12 task characteristics proved to be good predictors of said positive transfer