Publication: Generación de historias de vida usando técnicas de Deep Learning
Loading...
Official URL
Full text at PDC
Publication Date
2022-05-30
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
La memoria es un baúl donde almacenamos nuestros recuerdos más preciados dando sentido a nuestra propia identidad. No obstante, debido a ciertas enfermedades neurodegenerativas como es el Alzheimer, estos recuerdos pueden perderse. En este contexto, para paliar el deterioro de la memoria existen actualmente una serie de tratamientos no farmacológicos como son las terapias de reminiscencia que ayudan a no perder los recuerdos aun mantenidos. Dentro de este tipo de terapia destacan las historias de vida que están compuestas por los hechos más importantes acontecidos a una persona y son utilizadas tanto por el paciente como por el personal terapéutico para trabajar sus recuerdos y así retrasar al máximo su olvido.
La redacción de las historias de vida es un proceso complejo, ya que se suele realizar manualmente a partir de datos proporcionados por familiares o el propio paciente. Sin embargo, pese a los avances tecnológicos de los últimos tiempos en la Inteligencia Artificial, actualmente no existen apenas herramientas para tratar la redacción automática de la historia de vida a partir de los datos heterogéneos recogidos de la persona, hecho que facilitaría su elaboración.
Por todo ello este trabajo se centra en la investigación y desarrollo de un sistema de generación de historias de vida a partir de unos datos estructurados que representan las vivencias del paciente. El núcleo de este sistema reside en el componente modelador del lenguaje para el cual se considerarán aproximaciones de Deep Learning empleadas en sistemas similares de generación de lenguaje natural. Tras un estudio exhaustivo de distintos modelos de lenguaje y conjuntos de datos a utilizar, se ha seleccionado para este trabajo el actual modelo pre-entrenado T5 que será ajustado bajo el conjunto de datos WebNLG. Además, se han empleado algoritmos de clustering, entre otras técnicas, con la finalidad de inferir una estructura, orden y personalización al texto finalmente generado, mediante la manipulación de una serie de componentes en una interfaz desarrollada para la prueba del sistema. Por último, se ha realizado un estudio sobre la justificación de los resultados obtenidos al generar una historia de vida.
Memory is a boot where we store our most precious memories, giving meaning to our own identity. However, due to certain neurodegenerative diseases such as Alzheimer’s, these memories can be lost. In this context, to alleviate the deteriora tion of memory, there are currently a series of non-pharmacological treatments such as reminiscence therapies that help us not to lose the memories we still have. Within this type of therapy, life stories stand out, which are made up of the most important events that have happened to a person and are used by both the patient and the ther apeutic staff to work on their memories and thus delay their forgetfulness as much as possible. The writing of life histories is a complex process, as it is usually done manually from data provided by relatives or the patient him/herself. However, despite recent technological advances in Artificial Intelligence, there are currently hardly any tools to deal with the automatic writing of life histories from the heterogeneous data collected from the person, which would facilitate its elaboration. Therefore, this work focuses on the research and development of a system for the generation of life histories from structured data representing the patient’s experiences. The core of this system lies in the language modelling component, for which we will consider Deep Learning approaches used in similar natural language generation sys tems. After an exhaustive study of different language models and datasets to be used, the current pre-trained T5 model has been selected for this work and will be tuned un der the WebNLG dataset. Furthermore, clustering algorithms, among other techniques, have been employed in order to infer structure, order, and personalization to the text finally generated, by manipulating a series of components in an interface developed for the testing of the system. Finally, a study has been carried out on the justification ix of the results obtained by generating a life history.
Memory is a boot where we store our most precious memories, giving meaning to our own identity. However, due to certain neurodegenerative diseases such as Alzheimer’s, these memories can be lost. In this context, to alleviate the deteriora tion of memory, there are currently a series of non-pharmacological treatments such as reminiscence therapies that help us not to lose the memories we still have. Within this type of therapy, life stories stand out, which are made up of the most important events that have happened to a person and are used by both the patient and the ther apeutic staff to work on their memories and thus delay their forgetfulness as much as possible. The writing of life histories is a complex process, as it is usually done manually from data provided by relatives or the patient him/herself. However, despite recent technological advances in Artificial Intelligence, there are currently hardly any tools to deal with the automatic writing of life histories from the heterogeneous data collected from the person, which would facilitate its elaboration. Therefore, this work focuses on the research and development of a system for the generation of life histories from structured data representing the patient’s experiences. The core of this system lies in the language modelling component, for which we will consider Deep Learning approaches used in similar natural language generation sys tems. After an exhaustive study of different language models and datasets to be used, the current pre-trained T5 model has been selected for this work and will be tuned un der the WebNLG dataset. Furthermore, clustering algorithms, among other techniques, have been employed in order to infer structure, order, and personalization to the text finally generated, by manipulating a series of components in an interface developed for the testing of the system. Finally, a study has been carried out on the justification ix of the results obtained by generating a life history.
Description
Trabajo de Fin de Grado en Ingeniería Informática, Facultad de Informática UCM, Departamento de Ingeniería del Software e Inteligencia Artificial, Curso 2021/2022.
El código desarrollado durante la realización de este trabajo se encuentra disponible en https://github.com/NILGroup/TFG-2122-GeneracionHistoriasDeVida.