engleski

Too much of a good thing: unnecessary changes in translation students’ post-editing of machine translation

According to a recent EU-wide survey (ELIA et al. 2018), nearly 70% of translation companies and individual translators integrate machine translation (MT) in their workflows, with around 20% using it daily. This latest surge in MT use has been facilitated by considerable improvements in the quality of MT output in recent years, thanks to the development of neural machine translation (NMT). NMT systems use large amounts of data to train artificial neural networks to translate human language sentences. The improved output nevertheless contains many errors and requires human intervention for professional use. Such revision of MT output is known in the language industry as post-editing (PE). As a result of these trends in the translation profession, PE is increasingly seen as a skill which should be mastered by translation students, as well as translators already practising the profession. Previous research (Almeida & O’Brien 2010) has shown that untrained post-editors, even when they are experienced translators, tend to introduce unnecessary changes in MT output, which slows them down and thus defies the purpose of using MT in the translation workflow. Learning how to distinguish between necessary and unnecessary edits therefore seems to be an essential aspect of PE training. In a previous study (authors, forthcoming), we looked at the influence of time pressure and access to the source text (ST) on necessary edits in an NMT product. In a three-stage experiment, two groups of subjects – working with and without the ST respectively – post-edited the Croatian translation of an English ST produced by Google Translate, with varying degrees of time pressure. In the present study we analyse the same data, this time focusing on unnecessary edits. We distinguish among three types of such edits: those that, although unnecessary, improve the translation, those that consist in translation solutions that are neither better nor worse than original MT output, and those that in fact introduce an error where MT output was error-free. We compare data obtained in different conditions, investigating the impact of time pressure and access to the ST on the number and type of unnecessary edits. The results should contribute to our understanding of the post-editing process, thus proving relevant for the training of translators in PE tasks, especially from English into Croatian and related languages. References Almeida, Giselle de ; O'Brien, Sharon (2010) “Analysing Post-Editing Performance: Correlations with Years of Translation Experience.” In: Proceedings of the 14th annual conference of the European Association for Machine Translation, eds. Hansen, Viggo ; Yvon, François. URL: https://pdfs.semanticscholar.org/890e/c4a0b8a9d7b0f5e7bda3f7467f2cfec0cd4f.pdf. Elia ; EMT ; EUATC ; FIT Europe ; GALA ; LIND (2018) “The 2018 Language Industry Survey – Expectations and Concerns of the European Language Industry.“ URL: https://www.euatc.org/images/2018_Language_Industry_Survey_Report.pdf.

post-editing ; machine tranlsation ; unnecessary changes ; translator training

nije evidentirano