Can DeepL speak European Parliamentese? A case study into the error analysis of a report translated into Romanian
Can DeepL speak European Parliamentese? A case study into the error analysis of a report translated into Romanian

Summary/Abstract: Machine translation has gained much traction in the last decade, but research has been rather English-or Western-centric. This study looks at the errors produced by a public neural machine translation system (DeepL) when translating a European Parliament report from English into Romanian, an Eastern European language. More precisely, it focuses on a report published by the Committee on Women’s Rights. This paper puts forth a fine-grained report on the nature and severity of these errors while assigning a quality score and a pass/ fail threshold to the output. All errors are manually annotated using CATMA, a digital text markup tool. The error typology focuses on the grand categories of accuracy, fluency and fitness for purpose, using the Multidimensional Quality Metrics (MQM) framework as the basis. The qualitative analysis accompanied by examples offers translators and machine translation system developers an insight into the kind of errors they should expect and correct, as well as a template for calculating translation quality scores.

• Details
• Contents