Exploring lifelong learning in neural machine translation

Lifelong learning is a technique defined as the ability to learn newor maintain older knowledge over time. This factor is very important for humans, as it is one of the keys to the ability to learn by humans, which artificial intelligence tries to replicate. Additionally, it supposes the possibility...

Descripción completa

Detalles Bibliográficos
Autor: Guàrdia Fernàndez, Lluís
Tipo de recurso: tesis de maestría
Fecha de publicación:2021
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/360304
Acceso en línea:https://hdl.handle.net/2117/360304
Access Level:acceso abierto
Palabra clave:Active learning
Machine translating
Deep learning
lifelong learning
deep learning
quality estimation
machine translation
active learning
Aprenentatge actiu
Traducció automàtica
Aprenentatge profund
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Àrees temàtiques de la UPC::Enginyeria de la telecomunicació
Descripción
Sumario:Lifelong learning is a technique defined as the ability to learn newor maintain older knowledge over time. This factor is very important for humans, as it is one of the keys to the ability to learn by humans, which artificial intelligence tries to replicate. Additionally, it supposes the possibility for the systems to adapt to new inputs without the necessity to train the models from scratch every time. This technique, although being already applied on autonomous agents or machine learning for computer vision, has not been evaluated in the field of Machine Translation (MT). In the case of MT, currently, the majority of processes are still using sets of static data or traditional techniques which force to train the model only once without taking into account the possible variation of the language in the context. This project will be carried out on the context of a system implemented on the BEAT platform and thought for the evaluation of a lifelong learning task, which uses two sets of data: one for training and the other to apply the learning without the translations, having a simulated person to whom we can request the translations for the sentences we think are necessary. The objective is to suggest and analyze the usage of an active learning technique: Quality Estimation (QE), and the following comparison with the results obtained using random selection. In this project, we work with the language pairs of English-French and English-German. As results with QEwe achieve a score of 26.7 and 15.9 points for EN-FR and EN-DE, respectively, using a penalizing n-grams precisions BLEU score. These means an improvement of 0,5 for ENFR and 0,7 for EN-DE over the results obtained using random selection. In conclusion, the usage of lifelong learning in machine translation is feasible, although still is in an initial phase. As future possible actions over QE would be interesting to make a more extensive search of parameters or using an adaptive QE model.