Does simplification hypothesis still hold true for LLM-translated legal texts? A mixed-effects model approach

The simplification hypothesis in translation studies has been widely examined, especially in human translations, mainly through lexical simplification across modalities, language pairs and genres. Little is known, however, about the extra-lexical features of translations produced by large language models (LLMs), particularly in legal contexts. Moreover, most studies rely on Laviosa’s (1998a; 1998b) indicators, which lack genre specificity. This study addresses these gaps by testing whether simplification, through reduced linguistic variation and disambiguation, characterizes LLM-generated translations of legislative texts and by comparing them with human translations. A parallel and comparable corpus was compiled with Chinese source texts, their English human translations, LLM-generated translations and comparable English originals. Results show that LLM-generated translations exhibit simplification as both a T- and an S-universal in syntactic variation, global referential cohesion and certain sentential syntactic ambiguity indicators. Both human and LLM-generated translations reduce syntactic ambiguity, especially in the mean dependency distance for relative clause modifiers. However, human translators reduced lexical and syntactic variation more and better ensured global referential cohesion than LLMs, with implications for post-editing training, including attention to global referential cohesion and avoiding oversimplification of adverbial modifiers, as well as for LLM fine-tuning toward more human-like legal translations.