Abstract
Classical Chinese translation presents significant challenges: manual methods suffer from high costs and inconsistent quality, while both traditional machine translation and approaches relying solely on Large Language Models often fail to adequately capture intricate semantic nuances and cultural specificities. To overcome these limitations, this study proposes an LLM-driven multi-agent framework that decomposes translation into word-level interpretation, paragraph-level generation, and multi-dimensional review, integrating a specialized Key Word Interpretation Database, Retrieval-Augmented Generation, and iterative feedback. Experiments on The Records of the Grand Historian of China: The Hereditary Houses and the Biographies, Volume 7-10 show average improvements of 18.8-25.7% in BLEURT, BLEU-1, and METEOR over single-model baselines, with ∼ 12.7% reduction in score variance, indicating enhanced stability. Human evaluation confirms gains in fluency, adequacy, and cultural fidelity, particularly for weaker baselines. Ablation results reveal the indispensable roles of contextual coherence review, grammatical validation, and keyword interpretation, while efficiency analysis shows that compared with the framework without useful agents, the running time of the proposed method increases by 3.21 times, with the main contributing factor being the introduction of keyword interpretation. The framework excels in resolving polysemy, preserving cultural allusions, and improving semantic coherence. Beyond Classical Chinese, it offers a transferable blueprint for other historical or low-resource languages, supporting high-fidelity cultural heritage translation.