Contextual Translation of Halegannada to Hosagannada using Transformers

The preservation of linguistic heritage is challenged by the inaccessibility of older language forms, such as Halegannada (Old Kannada), which is rich in history and literature but difficult for modern speakers to comprehend. This research work addresses this gap by proposing a novel, two-phase Neural Machine Translation (NMT) system for contextual translation from Halegannada to Hosagannada (Modern Kannada). The proposed methodology leverages the multilingual mT5 transformer for initial lexical word construction, followed by contextual refinement using a locally deployed LLaMA 3.2 model to reconstruct poetic nuance and sociolinguistic richness. Evaluation on a meticulously hand-curated parallel corpus demonstrates the system’s superior performance, achieving a BLEU score of 48.5 against competitive baselines. This dual-phase approach ensures translations are not only accurate but also preserve the original text’s tone, context, and cultural integrity, significantly enhancing access to classical Kannada literature.