Método semi-supervisado para detectar, clasificar y anotar en un corpus de suicidio textos extraídos de entornos digitales
No Thumbnail Available
Date
2023-01-31
Journal Title
Journal ISSN
Volume Title
Publisher
Pontificia Universidad Católica del Perú
Abstract
La presente tesis doctoral, con un enfoque cualicuantitativo (mixto), se enmarca en la línea del análisis de sentimientos en redes sociales, forma parte del proyecto Life, que busca crear una plataforma integral para detectar y brindar apoyo especializado a usuarios de redes sociales que publican textos con contenido suicida. Por ello se desarrolló el Corpus Life para realizar experimentos con algoritmos de aprendizaje automático, mismo que originalmente constaba de 102 mensajes suicidas (71 textos en inglés y 31 textos en español), 70 de estas muestras Sin Riesgo y 32 con Riesgo. Pero debido al escaso número de muestras y al desbalance entre ellas, los resultados generados no eran confiables.
Por ello esta investigación tuvo como objetivo general desarrollar un método semi-supervisado para detectar, clasificar y anotar en el Corpus Life, textos extraídos de entornos digitales, con el fin de incrementar su número de anotaciones, mediante un proceso de evaluación automática de su calidad, previo a su inclusión o exclusión. Anotaciones que fueron evaluadas manualmente, utilizando para ello la medida de concordancia Cohen´s Kappa, con la participación de anotadores especializados quienes evaluaron los textos, alcanzando un nivel de acuerdo entre anotadores de 0,86, cercano al 0,78-0,81 de significancia estadística alcanzado automáticamente por medio del índice macro f1, con el método semi-supervisado. Lo que conllevo a alcanzar experimentos de un mayor grado de confiabilidad, por medio de un método estructurado con actividades, roles y procesos bien definidos y enlazados.
This doctoral thesis with a qualitative-quantitative (mixed) approach is part of the analysis of feelings in social networks that publish texts with suicidal content. For this reason, Corpus life was developed to carry out experiments with machine learning algorithms, which originally consisted of 102 suicide messages (71 texts in English and 31 texts in Spanish), 70 of these samples without risk and 32 with risk. But due to the small number of samples and the imbalance between them, the generated outcome was not reliable. Therefore, this research had the general objective of developing a semi-supervised method to detect, classify and annotate in the Corpus Life, texts extracted from digital environments, in order to increase their number of annotations, through a process of automatic assessments of their quality, prior to their inclusion or exclusion. Records which were tested manually, using the Cohen's Kappa concordance measure, with the participation of specialized annotators who evaluated the texts, reaching a level of agreement between annotators of 0.86, close to 0.78-0.81 of statistically significant reaching automatically by means of the f1 macro index, with the semi-supervised method. This led to achieving experiments with a higher degree of reliability, through a structured method with well-defined and linked activities, roles and processes.
This doctoral thesis with a qualitative-quantitative (mixed) approach is part of the analysis of feelings in social networks that publish texts with suicidal content. For this reason, Corpus life was developed to carry out experiments with machine learning algorithms, which originally consisted of 102 suicide messages (71 texts in English and 31 texts in Spanish), 70 of these samples without risk and 32 with risk. But due to the small number of samples and the imbalance between them, the generated outcome was not reliable. Therefore, this research had the general objective of developing a semi-supervised method to detect, classify and annotate in the Corpus Life, texts extracted from digital environments, in order to increase their number of annotations, through a process of automatic assessments of their quality, prior to their inclusion or exclusion. Records which were tested manually, using the Cohen's Kappa concordance measure, with the participation of specialized annotators who evaluated the texts, reaching a level of agreement between annotators of 0.86, close to 0.78-0.81 of statistically significant reaching automatically by means of the f1 macro index, with the semi-supervised method. This led to achieving experiments with a higher degree of reliability, through a structured method with well-defined and linked activities, roles and processes.
Description
Keywords
Aprendizaje automático (Inteligencia artificial), Procesamiento de datos, Suicidio
Citation
Collections
Endorsement
Review
Supplemented By
Referenced By
Creative Commons license
Except where otherwised noted, this item's license is described as info:eu-repo/semantics/openAccess