Neural and statistical machine translation: perception and productivity

The machine translation field has changed completely due to the many advances seen in neural machine translation (NMT) engines, especially in comparison with the results that were obtained with statistical machine translation (SMT) engines. So, it is necessary to review not only how MT is used but also how it is perceived by the end users, the translators. The main objective of this study is to determine the perception and productivity of a group of translators using SMT and NMT systems in terms of time and edit distance. Via the TAUS Dynamic Quality platform, ten professional translators first evaluated raw machine translation segments from two different texts- a user guide and a marketing text- proposed by the Microsoft Translation Engine (SMT) and Google Neural Machine Translation (NMT). Six of the ten translators subsequently post-edited two productivity tests to determine time and edit distance. The results show that translators perceive the NMT system as more productive because, according to their perception, it would take less time to post-edit and would mean fewer editions. However, when comparing these results with those obtained in productivity tests, although the edit distance was shorter when using the SMT engine than with the NTM, the post-editing time is much longer for the neural engine.