Title: Worlds Within Words: Translating Culture in Ancient Chinese Texts with Multi-Agent Coordination

Abstract:

While large language model (LLM)-driven machine translation has significantly enhanced cross-cultural dialogue, it continues to face difficulties when handling culture-loaded words (CLWs) found in classical Chinese literature. The core difficulty lies not merely in matching vocabulary but in determining the appropriate timing and method for explicating culture-specific knowledge for audiences who lack the necessary context. A strictly literal approach often retains surface-level forms but fails to convey deeper conceptual meanings, while excessive explanation can undermine the text’s conciseness and flow.

To resolve this dilemma, we frame the translation of CLWs as a task of selective explicitation. We introduce MACAT (Multi-Agent Culture-Aware Translation), a framework that dynamically detects culturally significant phrases and inserts brief, explanatory insights only when required. MACAT integrates a quality-aware reranking module to optimize candidate selection and employs a multi-round evaluation agent to judge translations based on five criteria: terminological accuracy, readability, fidelity, cultural preservation, and the effectiveness of cultural explicitation.

Our experiments, conducted on traditional Chinese medicine (TCM) classics and excerpts from the Analects, demonstrate that MACAT consistently surpasses both the underlying backbone model and standard machine translation baselines. These results were achieved under a unified GPT-5.4 evaluation setting, covering 100 TCM documents and a 20-chapter subset of the Analects.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC