Machine translation post-editing: What you need to know

Machine translation post-editing (MTPE) is the process in which a human linguist reviews and corrects a translation produced by machine translation or AI, so it reaches the quality the content needs. The machine produces the first draft quickly; the post-editor fixes errors, terminology and tone.

In this guide we cover the MTPE process, light vs. full post-editing, when it makes sense, common challenges, best practices, and a practical AI translation workflow.

What is machine translation post-editing?

Machine translation post-editing is a process in which human translators review and improve the output generated by machine translation systems. The role of MTPE is to enhance the overall quality and fluency of the translated content, ensuring that it meets the desired level of accuracy and naturalness.

The process explained

First, a machine translation system translates the source text into the target language. This is an automated process where algorithms analyze the input and generate a translation based on predefined patterns and models.

Then, a human translator (post-editor) reviews and edits the output to correct errors, improve coherence, and enhance the overall quality of the translation. The post-editor compares the machine-generated translation to the source text and makes necessary adjustments to ensure accuracy, fluency, and adherence to specific style or terminology requirements.

When do you need machine translation?

Businesses and translation professionals typically use automated translation when they need to translate a large amount of text quickly. MT can be much faster than human translation, so it is a good option for projects where you have a tight deadline.

MT and MTPE can also reduce the cost of translation. The two are far less expensive than human translation, which is great for businesses with a limited budget. Alas, you may need a translation that is accurate but not necessarily perfect. MTPE can nonetheless improve the accuracy of machine-translated text.

According to a survey from Unbabel, 39% of responders stated that they incorporate machine translation into their localization strategies. Among these users, 83% express satisfaction with the quality of the generated translations.

Light vs. full post-editing

The ISO 18587 standard for post-editing describes two levels:

• Light post-editing. You’ll get an understandable, accurate translation. The post-editor fixes errors of meaning, omissions and wrong terms, but doesn’t polish style. It suits internal content, support articles or high volumes where speed matters most.
• Full post-editing. You’ll get a translation that reads as if a human wrote it. The post-editor also fixes style, tone and flow. It suits customer-facing content: your product UI, marketing pages, anything with your brand on it.

Decide the level before you start, per content type, so post-editors know where to stop.

Post-editing AI translations

Large language models, LLMs in short, have really changed what post-editors work with. AI translations are usually more fluent than older machine translation, and they can follow instructions (tone, glossary terms, length limits, and more.)

But there’s also a downside: because the output reads well, errors are harder to spot. Just because the sentence is fluent doesn’t mean it can’t still have a wrong meaning, a missing negation, or an invented detail.

Quality estimation is useful in this aspect. QE is an automatic score for each translation that tells post-editors which strings to check first, instead of reading everything with the same attention.

When not to use machine translation

Do not use machine translation when you need a translation that is highly accurate and idiomatic. Machine translation is not as good at capturing the nuances of language as human translation, so it may not be the best option for documents that require a high level of precision.

If you require a translation that is confidential or sensitive, machine translation may not be a good translation option to consider. It can be a security risk, as the machine translation engine may store the source text and use it to train its models.

Finally, machine translation is not as good at producing creative or persuasive writing as human translation. Marketers may find that it is not the best option for marketing materials or other types of creative content.

Challenges of post-editing machine translation

Machine translation has come a long way in recent years, but it is still not perfect. There are a number of common mistakes that MT systems make, which can result in inaccurate or unnatural-sounding translations. These include things like:

• Omission of words or phrases. Sometimes, MT systems will omit words or phrases from the source text, which can change the meaning of the sentence.
• Addition of words or phrases. Conversely, MT systems can also add words or phrases to the source text that are not there.
• Errors in grammar and syntax. MT systems can also make mistakes in grammar and syntax, which can make the translation difficult to understand.
• Errors in meaning. Occasionally, MT systems will completely misinterpret the meaning of the source text.
• Literal translation. MT systems often translate words and phrases literally, without considering the context of the sentence. This can lead to translations that are nonsensical or grammatically incorrect.
• Difficulty with translating sarcasm and humor. MT systems often struggle with sarcasm and humor, as these rely on context and cues that are difficult for machines to understand.
• Difficulty with translating technical or specialized language. MT systems may not be able to translate technical or specialized language with the same accuracy as general text.
• Fluent but wrong output. It’s more common with AI translation. The sentence reads well, so the error is easy to miss.

Best practices for machine translation post-editing

Machine translation post-editing plays a crucial role in refining the output of machine translation systems to meet specific quality standards. Here are some best practices to consider when editing your translated content:

Understand the goal

Is the translation intended for informative, promotional, or technical purposes? You need to gain a nuanced understanding of the document’s context, industry-specific terminology, and the purpose of translation.

Employ high-quality source text

In order to avoid complex sentence structures, jargon, and technical terms that may be unfamiliar to MT engines, your source text needs to be easy to understand. You will also need to ensure the source text is free of grammatical errors and spelling mistakes, as these can also hinder MT accuracy. Maintain consistent formatting throughout the document, including font styles, paragraph breaks, and spacing.

Provide reference documents and glossaries

Provide relevant reference documents and glossaries for technical or specialized content to aid post-editors in understanding specialized terminology and concepts. Access to comprehensive reference materials can significantly improve the accuracy of translations in specialized areas.

Encourage collaboration among post-editors

Encourage post-editors to share their knowledge and best practices to improve the overall quality of translated text. You can leverage the combined expertise of post-editors to resolve complex translation issues and develop effective strategies.

Asses the quality

Scrutinize the machine-generated content for errors at various levels, including syntactic, semantic, and stylistic errors. You will need to address mistranslations, omissions, and structural issues. Also assess the natural flow of the translated text. Make sure that the sentence structures are not only grammatically correct but also maintain a coherent and natural progression, enhancing overall readability.

Ensure consistency

Use CAT tools to establish and maintain a consistent glossary of terms throughout the document. Check that the key terms are translated consistently, promoting a cohesive and professional final product.

Consider cultural sensitivity

To adapt the content to different cultural nuances, you will need to go beyond literal translation. This involves understanding idiomatic expressions, cultural references, and appropriately modifying the content to resonate with the target audience.

Use technology

Machine translation post-editing can be significantly enhanced with the use of various tools and technologies. Some commonly used tools include computer-assisted translation (CAT) tools, translation memory (TM) systems, terminology management tools, QA software, just to name a few. Or you can use a translation management platform like POEditor, that incorporates many of features of the aforementioned tools.

Evaluate MTPE performance regularly

Conduct regular evaluations of MTPE performance using metrics such as accuracy, fluency, and style to identify areas for improvement. One way to do that is to collect feedback from stakeholders and users of translated content. Afterward, use feedback and evaluation results to refine MTPE processes.

An MTPE workflow in POEditor

Now let’s see how a machine translation post-editing workflow looks in POEditor, step-by-step:

1. Translate automatically. Use AI translation with your own OpenAI, Anthropic (Claude), Google (Gemini), Azure AI or Amazon Bedrock account, or machine translation with Google Translate, DeepL or Azure AI Translator. A glossary and custom prompts keep terminology and tone on track.
2. Score the translations. AI Translation Quality Evaluation gives each translation a score, so you can see which ones need attention.
3. Fix the weak ones. Fix with AI suggests a revised translation for low-scoring strings. After a fix the string no longer has a score, so run the evaluation again to check the new version.
4. Review with people. Post-editors work on the strings that still score low, plus anything customer-facing. QA checks catch broken placeholders and inconsistent terms.
5. Mark what’s done. Proofread the translations that passed review. Proofreading doesn’t remove the Fuzzy flag, so clear it separately if your team uses it to mark strings for review.

You can automate these steps, including a re-evaluation after the fix, with a custom workflow (Premium plan and up).

Frequently asked questions

What is MTPE?

MTPE stands for machine translation post-editing: a human linguist reviews and corrects a translation produced by machine translation or AI, so it reaches the quality the content needs.

What is post-editing?

Post-editing is the human review of a machine-generated translation: fixing errors of meaning, omissions, terminology and, in full post-editing, style and tone.

What is the difference between machine translation and post-editing?

Machine translation is the automatic first draft. Post-editing is the human review that corrects and improves it.

What is the difference between light and full post-editing?

Light post-editing makes the translation accurate and understandable. Full post-editing also makes it read like natural, human-written text.

Is post-editing AI translation different from post-editing machine translation?

The task is the same, but AI translations are usually more fluent, so errors are subtler. Quality scores are used in AI translation with human review to catch these errors.

How do I decide what needs human review? 

Use quality scores and content type: review everything customer-facing and every low-scoring string; spot-check the rest.