TransLaw: A Large-Scale Dataset and Multi-Agent Benchmark Simulating Professional Translation of Hong Kong Case Law

TransLaw introduces a large-scale dataset comprising 344 professionally translated Hong Kong Court judgments, totaling 11,099 sentence pairs and 2.1 million tokens, addressing the need for parallel legal resources in English and Traditional Chinese. The framework decomposes translation tasks into multiple levels, utilizing a specialized legal glossary and Retrieval-Augmented Generation, and has been benchmarked against 13 LLMs, showing significant performance improvements over single-agent systems. This resource is crucial for AI practitioners focusing on legal translation, providing a robust dataset and a novel evaluation metric to enhance legal-semantic accuracy in machine translation.