Publication: Aplicación de Deep Learning a la Enseñanza de la escritura japonesa
Loading...
Identifiers
Publication date
2019-07
Defense date
2019-07-05
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
La enseñanza de la escritura de idiomas es un paso fundamental en el proceso
de aprendizaje de un idioma. Pese a esto, las aplicaciones más conocidas dedicadas
a los idiomas no se focalizan en este ámbito. Hay muchas aplicaciones pequeñas que
se especializan únicamente en la escritura de un idioma mediante el uso de plantillas caligráfcas correctoras del trazado. En este trabajo de fin de grado se propone
un nuevo método de enseñanza de la escritura: el uso de redes de neuronas convolucionales aplicadas al reconocimiento de imágenes de caracteres. Los caracteres
estudiados son los respectivos de los silabarios japoneses hiragana y katakana. Tras
una experimentación previa con otros dominios como CIFAR10, Dogs vs. Cats y
FERC, con el conjunto de datos de caracteres ETL Character Database se crean dos
modelos de predicción, uno por cada silabario, obteniendo un 99,23% de acierto en
el conjunto de validación en el modelo de katakana y un 99,61% en el de hiragana.
Posteriormente, se desarrolla una aplicación web dedicada al estudio y escritura de
caracteres japoneses en donde se integran dichos modelos, consiguiendo así un nuevo
método didáctico fiable para la escritura japonesa, libre de plantillas y que acepta
todo tipo de caligrafías.
The teaching of language writing is a fundamental step that is needed in the learning process of a language. Despite this, big known applications specialized in the teaching of languages do not focus on this matter. There are a lot of small apps that focus uniquely on the teaching of writing through the usage of calligraphic templates that often correct or even delete the drawn outline. This fnal degree project proposes a new writing learning method: the usage of convolutional neural networks applied to character images recognition. The studied characters are the ones from the two syllabaries of the japanese language: hiragana and katakana. After a previous experimentation with other domains such as CIFAR10, Dogs vs. Cats and FERC, the ETL Character Database is used to conform two prediction models, one for each syllabary, obtaining a 99.23% accuracy on the validation set for the katakana model and 99.61% for the hiragana one. Afterwards, a web application for studying and writing japanese characters is developed integrating those models, reaching a new trusted didactic method for writing japanese without any templates and that accepts any calligraphic style.
The teaching of language writing is a fundamental step that is needed in the learning process of a language. Despite this, big known applications specialized in the teaching of languages do not focus on this matter. There are a lot of small apps that focus uniquely on the teaching of writing through the usage of calligraphic templates that often correct or even delete the drawn outline. This fnal degree project proposes a new writing learning method: the usage of convolutional neural networks applied to character images recognition. The studied characters are the ones from the two syllabaries of the japanese language: hiragana and katakana. After a previous experimentation with other domains such as CIFAR10, Dogs vs. Cats and FERC, the ETL Character Database is used to conform two prediction models, one for each syllabary, obtaining a 99.23% accuracy on the validation set for the katakana model and 99.61% for the hiragana one. Afterwards, a web application for studying and writing japanese characters is developed integrating those models, reaching a new trusted didactic method for writing japanese without any templates and that accepts any calligraphic style.
Description
Keywords
Enseñanza de idiomas, Aprendizaje, Lengua japonesa, Escritura, Caligrafía, Redes de neuronas convolucionales, Deep Learning, Aprendizaje profundo