What is the recommended workflow to combine machine translation and human translations?

We have some languages that we would like to translate using Machine Translation, while for other languages (more relevant and important for us) we are willing to invest in human translation to guarantee more quality. How can I do that in Transifex and what is the best practices/workflow to accomplish it?

Hi Sandy,
In case you want to combine both, there are 2 different options you can choose to follow:
1- Turn on Auto-machine translation (available to PRO and Enterprise plans). You can run human reviews for the most important languages only.

2- Go to the editor, select the language you want to translate using Machine Translation, select ALL strings and trigger manually ā€œMachine Translationā€. For the other languages, you can order translations via Transifex Order Wizard, or inviting to the language teams for those languages your preferred vendor PM and request translations directly.

For any of them, make sure that you purchased the service with Google translate or Microsoft translator, and set the API key to enable Machine translation.

Let me know if you have further questions :slight_smile:
Kind regards,

I recommend that you do not publish machine translation without at least some proofreading. Even for languages where the results are generally good, you are bound to have some embarrassing grammar errors in your translation. For some language pairs, quality is so bad that itā€™s not even worth presenting it as a pre-translation to your localizers.

Machine translation is a supportive tool, it is not ready for creating production-quality copy.

@gunchleoc, Thanks for your comment!
Although we agree that this solution still might affect the translation quality, weā€™ve seen that the industry has gone more to this direction recently, and we are also looking for solutions where people can have better quality in terms of MT. We are working now on an integration with a QA tool that might help with a more robust review and fixes, for example, slight_smile:
Translators and reviewers will still have a great role here, but a more efficient way to work with MT and fix errors in an easier and quicker manner!

Yes, as I said, having it as a supportive tool is perfectly fine for those language pairs where the translation quality is sufficient that it makes the translatorsā€™ life easier.

The problem is that thereā€™s a common misconception out there that you can just use MT and slap it on your publication without post-editing. We are seeing this sort of sub-standard work crop up on everything from official signage to book publication.

OP sounded like they might want to publish for some languages without post-editing. These will probably be some lesser-used languages where MT quality will be pretty low. If you want to know more, check out the following articles:

When it comes to entire translations, the importance of a human translator cannot be overstated. A machine or programme that can translate with the same precision and nuance as a single or a group of human translators has yet to be developed.

A translatorā€™s particular understanding in corporate best practises or litigation is common. They can only translate the most important papers, as recognised by machines or summaries, because full translation is a lengthier, more thorough process.

Perhaps you find this helpful, @Sandyā€¦
Quality prediction is a key technology that has created a new workflow with MT. Itā€™s a technology that can estimate how MT performed on a segment and/or document level, so that users can anticipate post-editing effort.
With quality prediction, a new workflow is introduced: the hybrid translation process. There is a simple explanation of how hybrid technology works here: http://machinetranslate.org/hybrid-translation.

2 Likes

Hi Sandy,

You should take a look at machine translation quality prediction - AI that scores every machine translation segment.

With a quality score, you can prioritize your effort not just at the language level - a pretty crude lever - but also by the exact segment, according to whether or not it even requires human editing.

Traditionally, you had to choose between raw machine translation - zero quality checks - and machine translation post-editing - which isnā€™t significantly more efficient than full human translation from scratch. Between those two extremes is a massive gap is quality and price.

Now, with quality prediction, you can achieve the same final quality as machine translation post-editing up to 5x more efficiently.

And because itā€™s a score, you can also dial up the efficiency, if youā€™re willing to sacrifice a little bit of quality, but still want to catch the very bad errors.

Adam

Full disclosure:

Iā€™m the CEO and co-founder of ModelFront, the main provider of machine translation quality prediction.