Journal Paper Published

Abstract

Machine-translated language is now produced and disseminated on an unprecedented scale, emerging as a new variety of translational language alongside human-translated language. Yet its linguistic characteristics remain insufficiently understood. Focusing on discourse connectives, this study integrated exploratory correspondence analysis with confirmatory linear mixed-effects modeling to examine the discourse-level features of machine-translated language. The results showed that: (1) machine-translated language differed significantly from human-produced language, including both human translations and human-authored original texts, in academic abstracts and/or government documents, with machine-translated texts favoring additive connectives and human-produced texts favoring causal connectives, suggesting a contrast between safer, more general connective choices in machine outputs and more semantically specific discourse marking in human productions; (2) machine-translated and human-translated languages, both as varieties of translational language, used more additive connectives in government documents than human-authored original texts, which showed higher frequencies of adversative and temporal connectives, reflecting shared features of translational language; (3) across most connective categories and genres, no significant differences were found between NMT and LLM-based MT, suggesting convergence and relative stability in connective preferences despite differences in underlying AI paradigms; and (4) genre significantly moderated connective use across text varieties.

Keywords

connectives, correspondence analysis, genre, linear mixed-effects model, machine-translated language