LLMs for Machine Translation in Medium-Resourced Languages: From Transfer Abilities to the Impacts of Parallel Data

This thesis delves into the capabilities of decoder-only Large Language Models (LLMs) in the domain of machine translation (MT), with a focus on Iberian languages, particularly Catalan and Spanish. Through extensive experimentation using Falcon-7B and Aguila- 7B models, various aspects are investiga...

Descripción completa

Detalles Bibliográficos
Autor: Prats Cristià, Jaume
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
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:2117/422439
Acceso en línea:https://hdl.handle.net/2117/422439
Access Level:acceso abierto
Palabra clave:Machine translating
Deep learning (Machine learning)
machine translation
large language models
deep learning
Traducció automàtica
Aprenentatge profund
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial
Descripción
Sumario:This thesis delves into the capabilities of decoder-only Large Language Models (LLMs) in the domain of machine translation (MT), with a focus on Iberian languages, particularly Catalan and Spanish. Through extensive experimentation using Falcon-7B and Aguila- 7B models, various aspects are investigated, including fine-tuning strategies, parameter- efficient techniques, cross-lingual transfer abilities, and the impact of MT fine-tuning on natural language understanding (NLU) tasks. Key findings reveal the promising potential of decoder-only models for MT tasks. The study clarifies the impact of different quantities of parallel data on translation performance, emphasizing the significance of fine-tuning strategies. Additionally, it sheds light on the cross-lingual transfer abilities of decoder- only models in the related languages of the study. Moreover, insights into the transfer abilities between MT and NLU tasks are provided, highlighting associated constraints. Furthermore, the thesis identifies shortcomings in COMET metrics when handling off- target translation errors, higlighting the use of multiple metrics for result interpretation.