Publication: Búsquedas de respuestas con Deep Learning
Loading...
Files
Official URL
Full text at PDC
Publication Date
2018
Advisors (or tutors)
García Sevilla, Antonio F.
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
Citations
Exportar
Abstract
Esta memoria explica, en primer lugar, como funcionan sistemas de pregunta y respuesta existentes en la actualidad. En concreto, se estudian diversos sistemas relacionados con la medicina, debido que en un primer momento, en este trabajo se pretendía implementar un chat médico. A partir de conocer cómo funcionan los sistemas existentes y tener claro lo que se quiere conseguir, se sigue explicando los fundamentos técnico-teóricos más importantes del Deep Learning de cara a la implementación final del sistema de preguntas y respuestas. En concreto se hace mucho énfasis en los fundamentos del Machine Learning, los tipos diferentes de Redes Neuronales Recurrentes y como usarlas, y la arquitectura Encoder-Decoder. En el siguiente capítulo se explica la teoría del procesamiento del lenguaje natural. Este apartado de la memoria se explica en conjunto con Spacy, para darle también un enfoque más práctico. Al finalizar se explican los diferentes métodos de representación del lenguaje en vectores numéricos. Después empezará la segunda parte, y la más importante del TFG, que es como se han aplicado estos conceptos teóricos al sistema de preguntas y respuestas final. Se empieza por explicar lo que hace el código de manera secuencial y muy abstracta evadiéndose en todo momento del código en sí, para ahondar en por qué funciona de una forma más conceptual y teórica. A continuación, se hace mención a cada uno de los experimentos a los que se sometió el sistema y se comentan los resultados. Finalmente se procede a dar una opinión personal de por qué se han obtenido tales resultados y las soluciones que se le darían, haciendo hincapié en que se tienen que demostrar todos de manera empírica.
This report explains, first of all, how question answering systems currently exist. Specifically, various systems related to medicine are studied, because at first, in this work was intended to implement a medical chat. From knowing how the existing systems work and being aware about what we wanted to achieve, this document continues explaining the most important technical-theoretical foundations of Deep Learning for the final implementation of the question answering system. In particular, much emphasis is placed on the fundamentals of Machine Learning, the different types of Recurrent Neural Networks and how to use them, and the Encoder-Decoder architecture. The theory of natural language processing is explained in the next chapter. This section of the report is explained in conjunction with Spacy, to also give you a more practical approach. At the end, the different methods of representing the language in numerical vectors are explained. Then the second and most important part of the TFG begins, which explains how these theoretical concepts have been applied to the final question answering system. It begins by explaining what the code does in a very abstract and sequential manner, evading the code itself at all times, to focus into why it works in a more conceptual and theoretical way. Next, each of the experiments to which the system was underwent is mentioned and the results are discussed. Finally we proceed to give a personal opinion of why such results have been obtained and the solutions that would be given to it, emphasizing that they have to be demonstrated empirically.
This report explains, first of all, how question answering systems currently exist. Specifically, various systems related to medicine are studied, because at first, in this work was intended to implement a medical chat. From knowing how the existing systems work and being aware about what we wanted to achieve, this document continues explaining the most important technical-theoretical foundations of Deep Learning for the final implementation of the question answering system. In particular, much emphasis is placed on the fundamentals of Machine Learning, the different types of Recurrent Neural Networks and how to use them, and the Encoder-Decoder architecture. The theory of natural language processing is explained in the next chapter. This section of the report is explained in conjunction with Spacy, to also give you a more practical approach. At the end, the different methods of representing the language in numerical vectors are explained. Then the second and most important part of the TFG begins, which explains how these theoretical concepts have been applied to the final question answering system. It begins by explaining what the code does in a very abstract and sequential manner, evading the code itself at all times, to focus into why it works in a more conceptual and theoretical way. Next, each of the experiments to which the system was underwent is mentioned and the results are discussed. Finally we proceed to give a personal opinion of why such results have been obtained and the solutions that would be given to it, emphasizing that they have to be demonstrated empirically.
Description
Universidad Complutense, Facultad de Informática, curso 2017/2018