Beyond Fluency: Quantifying Grammatical Style Differences Between AI and Human Chinese-into-English Translation

This study investigates grammatical style differences between AI-generated translations (specifically GPT-4) and professional human translations (HT) in the Chinese-to-English direction, focusing on the concept of \"grammatical style inertia\"—a tendency towards reduced syntactic diversity and uniformity. Using a parallel corpus of 100,000 sentences from UN documents, we compared GPT-4 and HT outputs across multiple syntactic metrics (sentence length, subordination, passive voice, modality, POS/function word TTRs) and employed stylometric analysis (Burrows' Delta, hierarchical clustering, MDS) based on function word frequencies. Results consistently show that GPT-4 translations exhibit significantly shorter sentences, less subordination, fewer passive constructions and modals, and lower grammatical TTRs compared to HT. Stylometric clustering reveals a tight, homogeneous cluster for GPT-4 outputs, distinct from the broader, more variable HT cluster. These findings provide robust quantitative evidence for higher grammatical inertia in GPT-4 translations, suggesting that while fluent, its grammatical style is less rich and more uniform than professional human translation. We discuss the implications for translation quality assessment, the role of human translators, translator training, and AI development, contextualizing our findings with recent related studies and highlighting the unique contribution of quantifying grammatical inertia.