Computing for Multi-domain Neural Machine Translation (NMT) System for Sinhala, Tamil, and English

Principal Investigator: Nisansa de Silva

Co-Investigators: Surangika Ranathunga

We are trying to build high quality translation data sets for our low-resource languages (Sinhala and Tamil). However, to show that our manual data cleaning has made any effect, we need to train neural machine translation models with the original data and our new data. Also, the avoid the accusation of the models being less expressive just because of the size of the model we trained being too small, we need to do this with at least medium to large models.
The project proposes developing a multi-domain Neural Machine Translation (NMT) system for Sinhala, Tamil, and English to improve access to information across Sri Lanka’s official languages. It focuses on evaluating and improving existing parallel datasets, as well as creating new domain-specific corpora to ensure broad and reliable coverage. The final goal is to build a robust translation model, leveraging advanced architectures like mBART50, mT5, or NLLB, that performs well across multiple domains.

Objectives:

• Evaluate and improve Sinhala–Tamil–English parallel datasets by comparing neural machine translation performance trained on the original datasets versus manually cleaned and newly constructed datasets.
• Develop a robust multi-domain Neural Machine Translation (NMT) system for Sinhala, Tamil, and English using modern multilingual architectures (such as mBART50, mT5, or NLLB) to ensure reliable translation across diverse domains.
• Investigate the impact of model scale and architecture on translation quality by conducting controlled experiments across multilingual language models ranging from small (≈1B parameters) to large-scale models (up to 70B parameters).
• Analyze domain adaptation strategies for low-resource translation by training and evaluating both domain-specific and multi-domain models using curated datasets from multiple domains and language pairs.