Passive constructions in En/Fr human and machine translation

Abstract Passive constructions in English ( be+past participle ) and French ( être+past participle ) were inventoried on a text-per-text basis in a bi-directional French/English corpus. The corpus consisted of 70 (35 × 2) original source texts in each language, 70 (35 × 2) target texts produced by human translators, and the same number of target texts obtained by machine translation with DeepL Pro. Passive constructions were found to be about 70% more frequent in original English than in original French. In the corpora of translated texts, no large disparities were observed, with the exception of French texts translated from English by DeepL, which contained around 40% more passive constructions than original French texts. Correlations between frequencies of analogous passive constructions in pairs of source and target texts were statistically significant and stronger for machine translations than for human translations. Manual analysis of random samples of source/target pairs showed less regular correspondence between the two analogous constructions as counterparts for one another, and greater diversity among alternative translation counterparts in human translation compared with machine translation.