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

Researchers have released TransLaw, a multi-agent framework and dataset designed to simulate professional translation of Hong Kong case law from English into Traditional Chinese, a task mandated by the territory's Basic Law but long constrained by scarce parallel resources [1][2]. The system is built atop a new parallel corpus called HKCFA Judgment 97-22, which contains 344 professionally translated judgments comprising 11,099 sentence pairs and 2.1 million tokens, spanning rulings from 1997 through 2022 [1][2]. The work was posted to the arXiv preprint repository, an open-access platform that hosts scientific papers across disciplines including computer science and that, as of late 2024, receives roughly 24,000 submissions per month [6]. TransLaw breaks the translation task into three stages: word-level expression, sentence-level translation, and a multidimensional review [1][2]. The framework incorporates a specialized Hong Kong legal glossary database, Retrieval-Augmented Generation, and iterative feedback loops [1][2]. A four-dimensional expert review evaluates outputs for semantic alignment, terminology, citation accuracy, and judicial style [1][2]. Benchmarking across 13 open-source and commercial large language models — a class of machine learning models trained on vast text corpora for natural language tasks — showed that the multi-agent approach outperformed single-agent baselines on every model tested [1][2][8]. The system reached convergence within three iterations [1][2]. Ten certified legal translators then assessed the output using a proposed metric called Legal ACS. Their evaluation confirmed gains in legal-semantic accuracy, though the human experts still surpassed TransLaw in stylistic naturalness [1][2]. The dataset and benchmark code have been made publicly available on GitHub [1][2].