Hybrid translation

How to balance human quality and machine scale

Hybrid translation is used inside companies like Microsoft, Amazon and Unbabel, which invest in quality estimation research and development.

How can you use ModelFront translation risk prediction technology to get more scale than human translation and more quality than machine translation?

Default-human or default-machine?

We divide translation into default-human translation and default-machine translation.

Human translation can be sped up a little bit by using a translation memory or post-editing machine translations. That's often called "computer-assisted", but it's only a little bit faster than pure human translation.

Machine translation is highly scalable. It's 1000x cheaper than human translation, and nearly instant. It's integrated into top platforms, from user-facing text on marketplaces or social networks to backend operations like search or customer service.

Can we translate 2x, 10x or even 100x more than human translation?

Hybrid translation can be as close to machine translation as it is to human translation.

How it works

The hybrid translation workflow is smart and simple:

1. Machine translate

Get machine translations for all lines

2. Filter by risk

Get risk predictions for all the lines

3. Human translate

Have human translators post-edit the risky lines only

alt text

Example

Original
English
Machine translation
German
Predicted risk Human translation
German
...
...
...
Samsung Galaxy S7 edge Samsung Galaxy S7 Edge 1%
Various Colours & Grades Verschiedene Farben und Qualitäten 0%
Samsung Galaxy S7 EDGE Samsung Galaxy S7 EDGE S7 EDGE 96% Samsung Galaxy S7 EDGE
This Phone is Unlocked for GSM Networks Dieses Telefon ist für GSM-Netze freigeschaltet 1%
EMB - Phones, Accessories and More EMB - Telefone, Zubehör und mehr 0%
Charging/Data Cable Lade- / Datenkabel 2%
Wall Adapter Wandadapter 5%
Self Portrait Women's Abstract Triangle Lace Dress Self Portrait Women's Abstract Triangle Spitzenkleid 23%
A fairy-tale worthy Self Portrait party dress Ein märchenhaftes Selbstporträt-Kleid 84% Märchenhaftes Self Portrait Partykleid
We ship alll orders within 1 business day. Wir versenden alle Bestellungen innerhalb von 1 Werktag. 0%
For any question or concern, please contact us via eBay Bei Fragen oder Bedenken kontaktieren Sie uns bitte über eBay 1%
...
...
...

You can even send multiple machine translations, including translation memory matches, to let ModelFront select the best candidate.

Your key metric: machine translation quality

With ModelFront translation risk prediction, you'll be able to safely skip the majority of the segments that are machine translated perfectly. The single biggest success factor in hybrid translation is the quality of your machine translation.

What percent of your machine translations need post-editing today?

If that number is high, you should under stand why, and fix it. You can analyze the post-edits. You can break up your numbers by content type to see where it's working and not working. You can try custom machine translation, which has never been easier, or process improvements.

If that number is low, that's excellent. The efficiency is there, and now you can use hybrid translation to take advantage of it. You'll be able to safely skip the majority of the segments that are machine translated perfectly.

Tracking and improving machine translation quality is a good idea even if you are not using hybrid translation.


What's the best risk threshold?

If you want to hold quality equal, you can use the standard metrics. If you want to hold cost equal, you can just put the post-editing effort on the riskiest translations - a priority queue, no threshold.

For most platforms, it makes sense to apply application-specific business logic. For example, in a marketplace, titles have more impact on the final goal - sales - than descriptions do, so it's worthwhile to invest extra effort in translating them with quality. You may have different quality goals for some languages than for others.

How is it different than post-editing?

Reviewing and post-editing every machine translation is increasingly popular. However post-editing is only about twice as fast as pure human translation, and in many cases it's even slower.

With risk prediction, you can skip all human review for most lines - up to 10x faster than human translation.

How does it fit into our existing workflow?

ModelFront is a high-scale API that can be easily integrated from any platform. We work with enterprises, language service providers and toolmakers to help clients integrate translation risk prediction.

How can we adjust the balance of quality and scale?

ModelFront gives you the raw risk prediction, so you can decide the risk threshold for each content type, project, document or even line. We also provide custom models.