✨✍️ New: Read Automatic post-editing: the next step after quality prediction →
Automatic post-editing

Automatic post-editing — the next step after quality prediction

We're sharing a practical overview of automatic post-editing with you, a year after rolling it out in prod to most ModelFront customers.

To accelerate translation safely, they'd been using quality prediction – AI that learns which translations don't even require a human look.

But what about the millions of translations that require only small, repetitive, mechanical edits?

The natural next step is automatic post-editing (APE).

— Adam Bittlingmayer, co-founder, ModelFront, September 2025

Automatic post-editing (APE) is AI for fixing machine translations.

Why machine translation alone isn't enough

In the real world, the data shows human translators are forced to make millions of repetitive, mechanical manual edits to machine translation.

This is where automatic post-editing (APE) comes in – a separate LLM for generating fixes.

Why automatic post-editing alone isn't enough

The output of automatic post-editing is basically just yet another machine translation.

Now the real question becomes: Is it actually better than the original MT? Is it good enough to publish?

So automatic post-editing alone isn't solving much of a problem.

The missing piece

The leading translation buyers have shifted to safely automating millions of words at human quality.

The key is quality prediction – AI that learns which translations don't even require a human look, and which do.

✓ or ✗.

With automatic post-editing, buyers can automate even more.

Previously, they were skipping manual work on translations that humans would sign-off on as-is anyway.

Now, with APE, they can also skip manual work on translations that only need simple, repetitive fixes.

The new workflow

In a machine translation post-editing (MTPE) workflow, automatic post-editing is just part of a single AI step along with quality prediction.

Traditional manual post-editing workflow: MT → PE → 📄

New accelerated post-editing workflow: MT → QP → APE → QP → PE → 📄

  1. Machine translation (MT): Generates the initial translations.

  2. Quality prediction (QP): Checks which translations need a fix, and which don't – if not, it just skips manual human editing.

  3. Automatic post-editing (APE): Tries to fix the translation

  4. Quality prediction (post-APE): Checks again if the new translations (APE) needs a fix, or not – if not, it just skips manual human post-editing

  5. Human post-editing (PE): Only translations that still need a fix are sent to manual human post-editing.

Note that, under the hood, an automatic post-edit should always be re-checked with quality prediction.

Concrete value

In combination with quality prediction, automatic post-editing is AI that creates concrete value, by automating millions of words of translation at human quality.

  • Faster turnaround: Speed up delivery time
  • Cost savings: Save money
  • Scalability: Translate more content, or into more language markets, with the same resources

To be clear:

  • It's not for reducing post-editing effort in traditional manual post-editing workflows.
  • It's not for improving final quality in raw machine translation scenarios.

Sure, it'll help a bit, but there's not much concrete value.

Key takeaways

  • Automatic post-editing does not create value without quality prediction.

  • Automatic post-editing is already used by the translation buyers successfully, by combining it with quality prediction.

  • Together, automatic post-editing and quality prediction work like a human translator to take care of a big chunk of the work.

Did you find this helpful?

Share modelfront.com/automatic-post-editing (opens in a new tab) with your team and more folks it can help!

— Adam

FAQ

What kind of edits can APE do?

APE mainly handles edits that are relatively cosmetic – punctuation, whitespace, capitalization… But APE can also fix product names and terms, edit fuzzy matches and even do full-sentence rewrites, if they're predictable, which they often are.

The stuff that it's really a shame to waste human intelligence on.

But the key is quality prediction that keeps human quality. There is no limit to what can be generated.

Why do we need humans at all? Why can't APE do everything?

There will always be some input that is so bad, ambiguous or new that a proper quality prediction system will reject any translation, to trigger human intervention.

High-value content is fundamentally new – new products, new features, new inventions, new pandemics. There is repetition, which is why so much is automatable. But the first time a new term or new phrase is translated in a content stream, requires a human to define how it should be translated.

(Generic chat models like ChatGPT are also based on RLHF – Reinforcement Learning with Human Feedback. That is, they learn from human feedback, not just from the original training texts.)

Can APE be used for human translations?

Yes, APE is not just used for machine translations.

APE is used for many types of segments that would go to manual human post-editing – machine translations, fuzzy matches and exact matches.

It's even used for new human translations or post-edits that would go to a second manual human review.

Why doesn't machine translation do what APE does?

In theory, machine translation or other LLMs could generate better translations. We'd love that! But it never happened.

In practice, translation teams don't maintain custom machine translation for every content type and category, for every language – that would need to train hundreds or thousands of models, every few months. Many even just use generic models.

ModelFront models – both quality prediction and automatic post-editing – are aware of content type and category, and regular retraining is handled by the ModelFront team.

So our APE has an unfair advantage.