Neural MT and Human Post-editing: a Method to Improve Editorial Quality
Résumé
Machine translation (MT) has put more and more pressure on translators, especially since neural MT outperforms statistical MT. NMT provides better quality translations (more accurate and natural), and becomes closer to human translation. Yet, human translators must demonstrate their expertise and added value over such systems.
Our study is based on an on-going project in partnership with Presses Universitaires de Rennes (PUR, one of the major French publishers), Maison des Sciences de l’Homme en Bretagne (French Centre for Human Sciences) and the TRASILT team (Translation, Linguistic Engineering and Terminology) within LIDILE research unit (Language Linguistics and Teaching, EA 3874). It consists in devising a method for researchers that combines NMT (DeepL) and human post-editing to improve the quality of article metadata (abstracts, keywords, contents, etc.) from French to English in the editorial process of journals. The objective is to develop a methodology for translation that can be reproduced and transferred to other journals and disciplinary fields.
Based on the metadata of 75 articles published in 2017 in 4 PUR journals, it was decided to first compare the previously published English translation of these metadata with the NMT-generated translation of the same data. Second, an enriched version of the metadata (combining best translations from both outputs) will be post-edited and further improved by professional translators. Our goal is to determine the qualitative elements and limitations of each output (human vs. NMT) and design the most appropriate translation method. The method will then be tested on the 2020 issues of the 4 selected journals.