Machine Translation for Fassa Ladin

Overview

This repo contains my final project for Applied Natural Language Processing (University of Trento, a.y. 2023/24).

• Introduction
• Data
• Experiments
  • Preliminary Experiments
  • Finetuning
  • Evaluation (with Transfer learning across domains and Forgetting of previous knowledge)

Introduction

I construct the first Fassa Ladin-Italian-English parallel corpus, and train a machine translation model on it. More information can be found in the accompanying report.

🛠 English → Ladin demo

You can try translating text from English/Italian to Fassa Ladin using the model on Hugging Face Spaces 🦀

Data

The dataset draws from multiple resources in 5 different domains: literature, news, games, laws, and brochures. It is available in the data directory, either as a single file or split into train, validation, in-domain test, and out-of-domain test sets.

Experiments

Preliminary Experiments

Evaluate the performance of the pre-trained models.

Finetuning

Fine-tune the pre-trained models on the Fassa Ladin-Italian-English parallel corpus, with the two approaches: Pivot-based transfer learning and Multilingual translation.

Evaluation

Evaluate the models' performance, investigate Transfer learning across domains, and Forgetting of previous knowledge.