Exploration de la traduction automatique neuronale espagnol-français : Pour une traductologie de corpus appliquée à l’analyse des outils de traduction

Abstract

Exploring Spanish-French Neural Machine Translation: Towards a Corpus-Based Translation Studies Approach for the Analysis of Translation Tools

Neural machine translation (NMT) systems, based on large volumes of bilingual and monolingual data, represent a significant leap forward in the processing of linguistic data. They are capable of proposing a target text that is natural, fluid and idiomatic. The field of translation has gained much progress thanks to this technology wherein the working procedures are witnessing an important transformation. Automating the translation process requires not only to rethink the professional practices but also the objectives and the methods governing training in translation. In some professional settings, they are becoming a complementary tool for translation. The challenge is therefore to know the advantages and weaknesses of their use, which would make it possible to approach the post-editing phase more effectively. From a textometric and translation studies approach, this article proposes the exploration of a translation corpus composed of journalistic texts in Spanish and their translations into French by the generic NMT system DeepL. The exploration of the data comprises three stages: a textual approach, the analysis of the units selected by means of the cartographic representation of the correspondences and the return to the text, by examining the units in context. A sample of the most frequent lexemes in the corpus was qualitatively analysed using the MkAlign software in order to evaluate the output of the automatic information transfer. This is meant to verify the accuracy of the automatic translation process between two languages and the extent to which, if any, it causes information loss. The analysis indicates that, in most cases, the information content of the lexemes studied was correctly transferred into the target language. Furthermore, appropriate translation choices were revealed when processing anaphoric expressions. However, the target text contains transfer errors concerning some neologisms, proper names and parasynonyms.