Optimizing Chinese-to-English Translation Using Large Language Models

The advent of Large Language Models (LLMs) has significantly advanced Chinese-to-English translation tasks. However, the translation process remains challenging due to the substantial differences in syntax, semantics, and morphology between these two languages, despite notable achievements. This paper presents a comprehensive study on Chinese-to-English translation, evaluating the performance of various LLMs. We explore a range of open-source models, from 3.5 billion to 72 billion parameters, and OpenAI's latest models, across zero-shot, few-shot, and fine-tuned learning paradigms. Our analysis assesses translation quality using the COMET metric, reliability with the Translation Completeness Ratio (TCR), and efficiency via Characters per Second (CPS). The results highlight substantial trade-offs between model size, translation accuracy, and processing speed. Larger models tend to produce higher-quality translations, whereas smaller models offer greater efficiency. Fine-tuning significantly improves the performance of open-source LLMs, surpassing few-shot learning in both translation quality and processing speed. Proprietary models like GPT-4o exhibit consistent high performance without significant gains from fine-tuning. We emphasize the potential of fine-tuning with techniques like LoRA/QLoRA to optimize the balance between translation accuracy and computational efficiency, offering valuable insights for deploying LLMs in real-world translation scenarios.