Symmetrical Retrieval-Augmented Generation: A Risk-Averse Framework for Safety-Critical Translation of Biomedical Subject-matter

Large Language Models (LLMs) demonstrate high fluency in machine translation, yet their probabilistic nature can lead to hallucinations and terminological drift in dealing with safety-critical biomedical texts. To resolve this “fluency trap,” this study proposes a symmetrical Retrieval-Augmented Generation (RAG) framework powered by DeepSeek-V3. Functioning as a risk-management instrument, the architecture enforces a dual-constraint mechanism: a Lexical Stream applies rigid terminology injection to ensure high factual precision, while a Syntactic Stream retrieves structural templates from the authoritative New England Journal of Medicine (NEJM) corpus to preserve the professional style. Empirical evaluations on clinical trial protocols reveal that our framework significantly outperforms zero-shot GPT-4, achieving a near-perfect 99.2% Terminology Accuracy Rate (TAR) and superior semantic alignment. Ultimately, this study demonstrates that domain-specific, theory-driven architectures are more effective than larger model size, successfully transforming biomedical AI translation from probabilistic guessing into accurate, data-driven translation.