LinguaFuse: A Training-Agnostic, Context-Aware, Retrieval-Augmented Translation Framework for Low-Resource Living Languages

Running Head: LINGUAFUSE: LOW-RESOURCE CONTEXT-AWARE TRANSLATIONAuthor: Khan Tahsin Abrar Independent Researcher, Bangladesh https://orcid.org/0009-0009-4631-6768AbstractMost of the world’s 7,000+ living languages are underserved by modern machine translation (MT) systems due to exorbitant data and infrastructure requirements. Current best-practices rely on Reinforcement Learning from Human Feedback (RLHF) and extensive parallel corpora, making context-aware and domain-savvy MT unattainable for speakers of “low-resource” languages. This paper presents LinguaFuse: a novel, modular, training-agnostic MT framework that fuses retrieval-augmented lexicon lookup, rule-based grammar transfer, domain scoring, and lightweight reward modeling to render culturally and contextually apt translations. Unlike prior neural approaches, LinguaFuse is operational with zero LLM retraining, yet gains fluency with optional fine-tuning or system prompt adaptation. We demonstrate the feasibility and extensibility of LinguaFuse through a detailed Bangla implementation, cost analysis, and cross-linguistic validation proposals. The result is an auditable, human-in-the-loop, and democratized approach to high-quality MT, accessible to any individual or community, regardless of funding or computational resources.Keywords: Machine translation, RAG, grammar transfer, low-resource languages, Bangla, PEFT, human-in-the-loop, democratized NLP, HeX-HAG, Computer Science, Artificial Intelligence, Large Language Model (LLM)IntroductionThe rapid evolution of AI-powered language models has prompted a paradigm shift in access to global information. However, this access is starkly inequitable: only a handful of high-resource languages benefit from first-rate machine translation and linguistic technologies. The remainder, over 7,000 living languages, including Bangla (spoken by more than 230 million people), are relegated to inconsistent, often literal translations that fail in specialized, formal, or culturally nuanced domains (Ethnologue, 2025). Reasons include prohibitive reinforcement learning costs (often exceeding $25,000–45,000 USD per language: Epoch AI, 2024), lack of large parallel corpora, and local expertise drain.The core scientific question addressed herein is: