Thoughts on machine translation implementation

I am not a computer scientist, but I have been working with machine translation (MT) in the creative industries for almost a decade: testing it, measuring its quality and impact on productivity, drafting MT implementation plans, managing proofs of concept and training translators in post-editing.

I think we all agree by now that MT is here to stay. If anything, this past year has pushed the limits of existing localisation workflows, increased the speed of automation adoption and brought news of more MT implementation in subtitling. But why is it that many translators push back on MT, when it is supposed to be a tool designed to help them? “Nothing about me without me” is a disability activism slogan, and I would argue it stands true for translators as well. 

I was recently asked what my vision for the implementation of MT in translation workflows is and, after mulling on it, I think it boils down to this: MT implementation is not easy. It needs to be meticulously planned and executed if it is to be successful. There are two aspects that are crucial to guarantee success: the first has to do with the quality of the technology itself and the second with change management and how people interact with the technology. 

Everyone understands that MT quality needs to be as high as possible for its implementation to be successful. Yet MT quality is often judged by translators on the basis of the output from general purpose systems like Google Translate to translate e.g. subtitles, as that’s all they have at their disposal. Which probably feels like using a butter knife to cut meat. It goes without saying that customisation is important. Even better if you can customise not only to the domain, i.e. use subtitles to train a system designed to translate subtitles, but to your own data, to previous high-quality translations you have produced for the domain you work in, and then run a proof of concept to fine-tune the output further.

The other area that’s often overlooked is integration. Will users be able to reap the maximum benefits from a system built for them to use? Automatic quality estimation can block lower quality translations from view, and there are ways to trigger the MT to produce more accurate output. The hottest new trend in MT research is making use of metadata to increase MT quality. Whist some type of metadata can be used automatically within MT systems, e.g. the version of Spanish or Portuguese required in a given project, others rely on making use of data that has been created or collected upstream in the translation workflow, or on the translator providing such input. Metadata aside, the MT needs to be accessible by translators within their regular working environment, most typically via an API integration. But this is not enough; the user interface of translation editors needs to be adapted for MT use. Ergonomics are crucial to post-edit smoothly and with little fatigue. Pre-populating translation boxes with MT actually annoys many translators – just ask them! 

Which leads me to the next point: change management. If translators are to become MT adopters, MT quality is not enough, the entire process and workflow need to be communicated clearly and transparently to them, with relevant training provided and tight control processes in production. After all, they are the ones that will come up will all the great ideas about how to make such a tool more efficient, which makes their collaboration not only welcome but truly needed. With the right metrics in place, visible to the translators, one can make appropriate decisions regarding fair compensation in post-editing workflows, thus addressing one of the most important, if not the major point of apprehension among translators regarding the use of MT.

If you are a translator and you would like to give MT a go, here are a few questions you should ask yourself when you work with any given MT system. 

About the MT system:

  • Am I using a general purpose MT system or one specialised for my domain, e.g. subtitling?
  • Has the MT system been further customised with my or my client’s data?
  • How is MT integrated in the tool that I am using for my translation work? Has the UI been adapted to offer MT-related functionality that could help translators?
  • Are the term bases/glossaries I need to use integrated in the MT? 
  • Can I control the type of MT output provided to me, e.g. its length, or choose between different options?

The answers to these questions will tell you a great deal about the ease of the post-editing process and the usefulness of the MT system under evaluation as a tool in your line of work. 

About yourself:

  • What is my average speed when I translate?
  • What is my average speed when I post-edit the same type of text?
  • What/How many errors do I normally make in a file I am translating?
  • What/How many errors do I normally make when I am post-editing the same type of text?
  • How tired do I feel after post-editing for eight hours as compared to after translating the same type of text for eight hours?

The answers to these questions will tell you a lot about whether using the MT system under evaluation makes you more productive.

Have you found this article useful? Please contribute to the discussion by leaving a comment!

One Reply to “Thoughts on machine translation implementation”

Leave a Reply

Your email address will not be published.