A corpus-based search for machine translationese in terms of discourse coherence

Abstract

Earlier studies have corroborated that human translation exhibits unique linguistic features, usually referred to as translationese. However, research on machine translationese, in spite of some sparse efforts, is still in its infancy. By comparing machine translation with human translation and original target language texts, this study aims to investigate if machine translation has unique linguistic features of its own too, to what extent machine translations are different from human translations and target-language originals, and what characteristics are typical of machine translations. To this end, we collected a corpus containing English translations of modern Chinese literary texts produced by neural machine translation systems and human professional translators and comparable original texts in the target language. Based on the corpus, a quantitative study of discourse coherence was conducted by observing metrics in three dimensions borrowed from Coh-Metrix, including connectives, latent semantic analysis and the situation/mental model. The results support the existence of translationese in both human and machine translations when they are compared with original texts. However, machine translationese is not the same as human translationese in some metrics of discourse coherence. Additionally, machine translation systems, such as Google and DeepL, when compared with each other, show unique features in some coherence metrics, although on the whole they are not significantly different from each other in those coherence metrics.