The industrial process for quality machine translation

temperature-gauge-in-industrial-processMachine translation (MT) is not a tool. Machine translation is an industrial process. Selecting the right MT engine is just one part of a process that, if correctly managed, is capable of lowering translation costs, increasing productivity and even improving quality and consistency. Quality MT output is critical because it determines productivity. With quality as the goal, the question is not so much what engine to choose but what engine and what process to leverage human input will give the best results.

If you treat MT as a process rather than as a single engine, quality machine translation IS possible.

Lexcelera, founded in 1986, has been deploying Machine Translation (MT) systems for customer content since 2007, localizing on a variety of engines in over a dozen languages. We have built engines for the ‘easy’ languages, French and Spanish, as well as for Italian and Dutch, German, Polish, Swedish, Japanese, Chinese, Arabic, and more. We have used a variety of commercial and non-commercial engines, rules-based (RBMT), statistical (SMT) and hybrid.
In 2012 Lexcelera opened LexWorks to be the North American bridge to MT services such as consulting, creating engines and post-editing.

Over the years we have delivered major translation projects using machine translation to shave off 50% of the time and 30% of the cost. (Our customers have achieved a positive Return on Investment – or ROI – from the first project, including engine training costs.) We have also delivered documentation, OLH (online-help) and courseware translations that were more highly rated for quality than their fully human counterparts. We have turned around 90,000 word translations overnight in raw MT so that our customers could respond to calls for tender more quickly. Working with Moses inventor Philipp Koehn and the University of Edinburgh, as well as other partners such as Acrolinx and Symantec, we have been awarded a major EU grant for a ground-breaking research project.

And what has all this taught us?

We have come to realize that machine translation is not a tool. Machine translation is a process. While much debate centers around the engine, whether rules-based or statistical, we have found that selecting the engine (or engines) is just one part of an industrial MT process – a process that, if correctly managed, is capable of lowering translation costs, increasing productivity and even improving quality and consistency.

To reach this goal, we believe that eight stages need to be in the MT process.

Pre-production

  • Consultation
  • Content
  • Customization

Production

  • Piloting
  • Processing
  • Post-Editing

Post-production

  • Metrics
  • Maintenance

The last of these stages, maintenance, is the virtuous circle that links post-editing review feedback to further customizations and real-time improvements in quality. In this process, machine translation technology is augmented by humans as well as by automated quality controls on both sides of the production process (automatic pre- and post-editing). Ongoing improvements come from correcting issues flagged by the post-editors.

Quality machine translation is possible, as long as it is part of a solid industrial process to maximize both human input and all automated tools in our toolbox.