An Attention-Driven Multilingual Ethnic Language Translator (MELT) for the Marma Language: A Comparative Analysis of LLaMa-3 & NLLB-200

Abstract

Bangladesh is a country of many languages and cultures where ethnic groups reside collaboratively in harmony and tranquility. But there remains a language barrier among these groups and the locals which hinders ethnographic research and development. This study addresses such digital divide affecting the Marma language, an indigenous language spoken by approximately 200,000 people in Bangladesh, but lacks standardized digital resources and translation tools. In this research, we developed an Attention-Driven Multilingual Ethnic Language Translator (MELT) that facilitates bidirectional translation between English, Bangla, and Marma. Using field data and secondary sources, a new ethnographically validated parallel corpus of over 12,000 sentences was created. To reduce data sparsity, transliteration and curriculum learning were used. The NLLB-200 multilingual model, and an improved LLaMA-3 (8B) model were the two architectures that were benchmarked. In low-resource settings, our experimental results show that the refined LLaMA-3 significantly outperforms multilingual NMT models and conventional neural baselines, especially when it comes to managing Marma script variability and semantic consistency, supporting linguistic inclusion and digital accessibility. Word Accuracy, Character Accuracy, Word and Character Error Rates, and BLEU-4 scores have been used to evaluate the performance of these models. According to our evaluation, LLaMa-3 outperformed the other models with a BLEU-4 score of 71.04, Word Accuracy Score of 64.3%, WER of 0.22, and a CER of 0.18.