How dependency-based syntactic complexity shapes post-editing of LLM-generated translations

Abstract This study examines how dependency-based syntactic complexity shapes English-to-Chinese post-editing of GPT-4 output. We address three questions: (RQ1) whether syntactic complexity on the source-text and GPT-output sides predicts post-editing quality; (RQ2) whether translator expertise moderates syntactic complexity effects; and (RQ3) whether task conditions affect performance and moderate syntactic complexity effects. We collected data from 46 participants (30 students, 16 professionals) post-editing in Trados Studio. We operationalized complexity with dependency-based metrics grounded in Incomplete Dependency Theory and Dependency Locality Theory, complemented by English–Chinese–specific indices (Left-Embeddedness; Nested Noun Distance). Results show that complexity in both source texts and GPT-4 output predicts post-editing errors (RQ1); experts are less sensitive to rising complexity than students, especially at higher complexity levels (RQ2); termbase access reduces overall and terminology errors, and the expert advantage in handling complexity is larger under Light post-editing but attenuated under Full post-editing (RQ3).