Enhancing Low Resource Machine Translation Through Multilingual Generative Models and Synthetic Data Augmentation

Generative Artificial Intelligence (GenAI) has achieved remarkable success in machine translation for high-resource languages; however, low-resource languages continue to suffer from poor translation quality due to limited parallel corpora, complex morphology, and cultural contextualization challenges.This research presents an enhanced generative translation framework tailored for low-resource languages by combining multilingual data augmentation, transformerbased architectures, and culturally aware optimization strategies.The proposed system leverages synthetic dataset generation through back-translation and large language model augmentation, followed by fine-tuning of mT5, mBART, and LLaMA models using parameter-efficient techniques such as LoRA and adapter layers.To reduce hallucinations and improve semantic fidelity, constraint-based decoding and reinforcement learning from human feedback are incorporated.Experimental evaluation conducted on multiple low-resource language pairs demonstrates significant performance gains over baseline models, achieving an average improvement of 8.7% in BLEU score, 6.3% in COMET, and 7.9% in BERTScore.Additionally, hallucination rates are reduced by approximately 35%, while human evaluation indicates improved grammatical correctness and cultural appropriateness in over 82% of translated samples.These results validate the effectiveness of the proposed framework in delivering accurate, context-aware translations for low-resource languages.The study contributes a scalable and deployable GenAI-based translation solution that promotes digital inclusion and supports linguistic preservation in multilingual ecosystems.