Low-resource-Machine-Translation

This repository contains the code for the project relative to the course Deep Natural Language Processing. The goal of the project is to replicate the experiments performed by Dabre et al. on low-resource machine translation. In particular, starting from a machine translation model pretrained on a large dataset, we finetune it on a low-resource language.

Implementation details

The initial model chosen for the task is MarianMT, a transformer-based model pretrained on a large English-Chinese corpus. The model is finetuned on three low-resource languages from the ALT dataset (Vietnamese, Indonesian and Filipino). The finetuning is performed using the Huggingface 🤗 Transformers library.