Reinforcement Learning for the Computational Interpretation of Classical Medical Heritage Texts

Traditional Chinese Medicine (TCM) classics are a major form of intangible heritage, preserving historically layered medical knowledge, diagnostic logic, and therapeutic epistemologies. For heritage-text digitisation, interpretive fidelity and epistemological continuity are as critical as linguistic fluency. We present R1-TCM-Translator, a heritage-oriented framework for ancient-to-modern Chinese medical translation that combines multi-objective reinforcement learning (GRPO) with a structured six-step reasoning process to make cultural-epistemic reconstruction explicit and auditable. Experiments on a philologically curated parallel corpus of 15,387 sentence pairs from eight representative TCM classics show consistent improvements over supervised fine-tuning baselines and strong general-purpose large models. R1-TCM-Translator-8B demonstrates consistent gains on lexical-alignment and semantic-consistency metrics, specifically BLEU and COMET, indicating improved cross-text interpretive stability rather than metric-specific optimisation alone. Fine-grained analyses further show an approximately 23-percentage-point increase in specialised terminology accuracy and improved consistency in interpreting complex pathogenesis-related semantics. These findings suggest that reward-guided structured reasoning can improve epistemic fidelity in digitised medical heritage archives beyond surface-form translation quality. By embedding semantic-fidelity objectives directly into optimisation, the framework operationalises heritage-aware translation as an auditable alignment process rather than a purely generative task. While doctrine-dense passages and composite interpretive risks remain challenging, the framework provides a reproducible computational pathway for digital preservation, knowledge modelling, and digital-humanities research on medical heritage texts.