MT quality metrics seem to study each segment in isolation, while an assessment of whether a professional translation is fit for purpose will normally consider the text as a whole. Some proponents of NMT (Neural Machine Translation) are claiming that professional translators are rating NMT output as “near-human”, but this is still at segment level.

  • To what extent will it be possible to refine the quality metrics used in assessing MT so that they take a holistic view?
  • Is it not the case that while a good professional translation is greater than the sum of its parts, any form of MT is still just that – a collection of parts stitched together to produce a less than cohesive whole?

  • Published: 3 ans ago on 11 janvier 2018


  1. Aljoscha Burchardt dit :

    As explained in my previous post, the automatic measures are in fact holistic in that they compare the MT output with a human-translated reference document. But, the comparison is so shallow that aspects of cohesion cannot be checked. The more analytic metrics often need human involvement which makes it possible to include a holistic view. I don’t see how this can be automated with current technology.