Low-Resource Myanmar-Dawei Machine Translation using Fine-Tuned Neural Models

In the current machine translation (MT) landscape, the Transformer architecture remains the gold standard, particularly for high-resource language pairs. However, low-resource language pairs—especially those involving regional dialects like Myanmar and Dawei—continue to pose significant challenges due to the scarcity of parallel corpora and the linguistic diversity they exhibit. This limitation highlights the necessity of fine-tuning models and optimizing hyperparameters to enhance translation performance, particularly for the Myanmar-Dawei language pair. This research explores the fine-tuning of pre-trained Opus-MT models on a Myanmar-Dawei parallel corpus for bidirectional translation tasks. We compare a baseline model, Transformer-based NMT, to fine-tuned Opus-MT models. The fine-tuned Opus-MT model achieved impressive BLEU score of 56.49 for Dawei to Myanmar and 45.40 for Myanmar to Dawei translations.This improvement demonstrates the effectiveness of leveraging pre-trained models for low-resource and dialectal translation tasks, contributing to the preservation of linguistic diversity and enhancing digital communication for underrepresented dialects such as Dawei. These findings underscore the importance of employing specialized fine-tuning of pre-trained models for low-resource language pairs and suggest promising directions for further optimization.