Uncovering Machine Translationese: an experiment on 4 MT systems for English-French translations

The aim of this presentation is to discuss the linguistic features of machine-translated texts in comparison with original texts in order to uncover what has been called “machine translationese” (e.g. Daems et al. 2017). Using a corpus-based statistical approach, namely, the Principal Component Analysis technique, 4 MT systems have been investigated for English to French translations of press texts: 1 Statistical MT (SMT) and 3 Neural MT (NMT) systems, namely DeepL, Google Translate, and the European Commission’s eTranslation MT tool, in both its SMT and NMT versions. In particular, to complement a previous study on language-specific features (e.g. derived adverbs, existential constructions, coordinator et, preposition avec, see Loock 2018), a series of language-independent linguistic features were extracted for each text, ranging from superficial text characteristics such as the average word and sentence length, to frequencies of closed-class lexical categories and measures of lexical diversity.The final aim is to uncover linguistic features in MT texts that clearly deviate from the expected norms in original French.