A multi agent classical Chinese translation method based on large language models

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Digital Humanities

Background:

• Classical Chinese translation faces challenges with manual methods and current machine translation (MT) approaches.
• Existing MT and Large Language Models (LLMs) struggle with semantic nuances and cultural specifics in Classical Chinese.

Purpose of the Study:

• To develop an LLM-driven multi-agent framework to enhance Classical Chinese translation quality.
• To address limitations in semantic accuracy, cultural fidelity, and consistency in automated translation.

Main Methods:

• A multi-agent framework decomposing translation into word interpretation, paragraph generation, and review.
• Integration of a Key Word Interpretation Database, Retrieval-Augmented Generation, and iterative feedback loops.
• Utilizing LLMs for nuanced interpretation and generation, supported by specialized databases and review agents.

Main Results:

• Achieved 18.8-25.7% improvement in BLEURT, BLEU-1, and METEOR scores over single-model baselines.
• Demonstrated a 12.7% reduction in score variance, indicating enhanced translation stability.
• Human evaluations confirmed superior fluency, adequacy, and cultural fidelity, especially compared to weaker baselines.

Conclusions:

• The proposed LLM-driven multi-agent framework significantly advances Classical Chinese translation.
• The framework effectively handles polysemy, cultural allusions, and semantic coherence.
• This approach provides a transferable model for translating other historical or low-resource languages, preserving cultural heritage.