Bridging the Linguistic Divide: A Comprehensive Review of Methods and Challenges in Low-Resource Machine Translation

Abstract

Large language models (LLMs) have become useful in high-resource languages due to the rapid advancement of natural language processing (NLP), yet the vast majority of the world’s languages remain underserved because of a severe lack of data. This paper develops a complete taxonomy of machine translation (MT) methods for low-resource languages (LRLs) and critically evaluates their construction. Rule-based (RBMT) and statistical machine translation (SMT) remain essential for digitally dark languages such as Tulu and Kangri, while transformer based frameworks and parameter efficient fine tuning (PEFT) demonstrate fluency in mid resource scenarios. We also evaluate data-centric interventions, balancing the advantages of parallel corpus mining against the risks of synthetic data, which can distort semantic polarity in sensitive domains. Our synthesis identifies crossover points where statistical methods outperform neural approaches and integrates findings from dynamic language identification, hybrid transfer learning, and retrieval augmented generation (RAG). The paper concludes with a resource-tier recommendation table and proposes future directions including neuro symbolic hybrid architectures, unsupervised corpus mining, and culturally grounded evaluation metrics.