Machine Translation: Friend or Foe? Translator Attitude, Process, Productivity and Quality in Human Translation and Post-editing Machine Translation

Publication type
Publication status
Daems, J, Tondeleir, L., & Macken, L.
View in Biblio
(externe link)


As the need for translation increases, the usage of machine translation (MT) increases accordingly. While many experienced translators distrust MT output, studies have shown an increase in productivity and quality when contrasting the post-editing (PE) of machine translation with human translation. We believe that a better understanding of the PE process and product is key to the general acceptance of MT+PE by the translation community.

We present some of the findings of our research on the differences between human translation and MT followed by PE. Novice translators (Master’s students in translation) were asked to perform both a translation task and a post-editing task, from English into Dutch. The selected texts were newspaper articles taken from the Dutch Parallel Corpus. Translation and post-editing processes were recorded using PET (a post-editing tool developed by Wilker Aziz and Lucia Specia). As the attitude of translators towards MT is of paramount importance for the development of the translation industry, participants had to provide feedback on the translation or post-editing process after each sentence. In the feedback for human translation, participants were asked to estimate the level of difficulty and to report on any translation issues. Feedback forms for PE tasks contained questions about the quality of the MT as well as questions to assess MT output difficulties. Translators and post-editors were also asked to indicate for which language items they consulted external resources and which ones they used. Quality was assessed for both adequacy and acceptability.

We found a significant correlation between subjective difficulty scores and time spent on the translation. Results further show that MT followed by PE is faster than HT and of comparable quality to HT. There was, however, a notable difference in the type of errors detected for both types of translation. For example, almost ten per cent of all PE-problems consisted of word sense shifts, compared to four per cent of word sense shifts for HT.