Formal divergence and legal relation preservation in Hong Kong judgment translation

Abstract

Abstract Legal English secures precision through dense qualificational syntax, and the evaluation of machine-translated legal texts often reads lower structural complexity as weaker legal precision. Whether formal simplification actually entails the loss of legally relevant relations has rarely been tested. This study examines the question in 57 Chinese-language Hong Kong judgments, comparing their official English translations with DeepSeek-V3.2 and ChatGPT output generated under a minimal prompt. Relative to the official translations, both model translations are shorter, lexically denser and markedly lower in subordination and complex nominal density. Their higher all-word lexical diversity lies in function words and has no counterpart in content words. A prespecified relation-level analysis then traced 192 source instances of qualification and scope marking into all three versions. Each version preserved the traced relations in over 97 % of its evaluable renderings, 98.06 % when the three versions are pooled. Of the eleven shifts, two occur in the official translations, five in DeepSeek and four in ChatGPT. The judgments with the greatest complex nominal divergence showed no larger model excess of shifts. The shifts that altered the reading of a judgment included changes in quantifier force, party coverage, attribution and restrictive scope, and the aggregate profile did not identify where they occurred. The findings separate formal simplification from relational loss. Legal precision in translated judgments is better evaluated through relational fidelity than through structural density, while corpus profiles remain appropriate for describing how production conditions package legal meaning.

The authors' abstract, as published at the source. International Journal of Legal Discourse, 2026 · DOI ↗