Depression classification model on Twitter using BERT

• Autores: Aleman-Zambrano, Guillermo José, Del Carpio-Lazo, Marvik Irzovic, Mendiguri-Chávez, Daniel Gustavo, Vilchez-Silva, Daniela Carolina, Tejada Toledo, Franco Eduardo
• Formato: artículo
• Fecha de Publicación: 2023
• Institución: Universidad La Salle
• Repositorio: Revistas - Universidad La Salle
• Lenguaje: español
• OAI Identifier: oai:ojs.revistas.ulasalle.edu.pe:article/89
• Enlace del recurso: https://revistas.ulasalle.edu.pe/innosoft/article/view/89; https://doi.org/10.48168/innosoft.s12.a89; https://purl.org/42411/s12/a89; https://n2t.net/ark:/42411/s12/a89
• Nivel de acceso: acceso abierto
• Materia: Depression classification; text classification; natural language processing; BERT; social networks; Clasificación de depresión; clasificación de texto; procesamiento de lenguaje natural; redes sociales

Today there are many signs of depression, as well as many suicide attempts caused by this emotional disorder, and this is reflected mostly on social networks, mainly on Twitter. For this reason, it is important for specialists and organizations seeking to safeguard people's lives to use software tools to address this problem. For this, in this work a web tool called "UBDevs-Depression-Classifier" is proposed,  that allows you to automatically obtain and classify tweets for a specific topic. A greater emphasis was placed on tweets related to COVID-19in the years 2020-2021 the world experienced a pandemic that increased cases of depression in many places. This research proposal focuses on the use of a model based on NLP (Natural Language Processing) for the classification of Tweets in order to find those that incite depression or imply that users are in a bad mood, all this in order to maintain the mental and physical health of the users of this platform. There are several models that are used as a basis for NLP projects, however, at present BERT has proven to be one of the most efficient, so we selected it for the development of our proposal. To evaluate the efficiency of the project we applied the F1 metric obtaining a value of 0.8806, a quite acceptable result with respect to a textual classification.