Making Machine Translation Sound Human
Machine translation is increasingly convincing, but it still needs human intervention to make it sound, well, human. Here's how I do that with post-editing.
What is Post-Editing?
Post-editing (often referred to as machine translation post-editing, MTPE or PEMT) involves reviewing a translation generated by a machine like Google Translate large language models like ChatGPT and refining it to improve accuracy, clarity and tone. It's different from revising, reviewing or proofreading in that the aim is not always to produce a perfect text. A post-edit might involve just tweaking a few awkward phrases so the text makes sense, or it could mean a significant rewrite to make the translation read as if it were written by a human.
The rise of AI tools like ChatGPT, Gemini and Claude has dramatically improved machine translation quality since the first attempts back in 1933. But they still can't match human translation, and this is where post-editing comes in.
Post-editing isn't always the best option, especially for creative or highly nuanced content, which usually benefits from being translated from scratch. But when time and budget are the top priorities, having a human post-edit a machine translation can be a fast, cost-effective solution.
What's the difference between machine translation and AI translation?
Neural Machine Translation systems, like DeepL and Google Translate, are explicitly trained for translation and tend to make predictable errors, such as awkward phrasing or overly literal translations, which are usually easy to spot and fix. LLMs like ChatGPT, by contrast, are trained on vast amounts of text to generate human-like language across a wide range of tasks, with translation being just one of them.
While LLMs often handle context better than traditional MT systems and generally produce more fluent, natural-sounding output, they still struggle with nuance, idioms, cultural references and implied meaning. The result is often text that, at first glance, looks polished but which, on closer inspection, doesn't quite capture the original message or tone.
More worryingly, LLMs can introduce subtle, harder-to-spot errors, like changing meanings, inventing details or leaving out information that was in the source. This means a ChatGPT translation, while more fluent, might actually need more careful editing than one produced by a tool like Google Translate. There’s a great article about the limits of automated translation, and we still need human translators here.
Although both technologies rely on neural networks, they behave differently, and each calls for its own type of post-editing. Most post-editors offer three levels of post-editing to suit different needs. These edits can also be applied to Dutch-to-English translations written by non-native English speakers.
Levels of post-editing
Light post-edit – Minimal intervention to ensure the text is comprehensible. Best suited for internal or short-lived content. Typically only needed for machine-translated text rather than output from LLMs.
Fixes major spelling, grammar and punctuation errors
Ensures everything is translated and removes obvious mistranslations
Does not include stylistic improvements or terminology consistency checks
Standard post-edit – A more polished version for short-term or external content. Usually recommended for machine translation, but LLM output will often also benefit from a standard post-edit. Includes all the checks in a light post-edit, plus:
Fixes major and minor spelling, grammar and punctuation errors
Makes minor stylistic changes to improve readability
Ensures key terminology is used correctly (if a glossary is provided)
Full post-edit – A high-quality, human-like translation. Recommended for LLM-generated content and any text intended for long-term use. Includes all the checks in a light and standard post-edit, plus:
Refines structure, tone, flow and localisation
Rewrites sections for clarity and natural style
Ensures consistency and accuracy across the text
Includes a full manual check for spelling and grammar
Note that post-editing does not usually include a full proofread. If the text is for external publication or long-term use, you should still have it proofread by someone else. However, even with this additional step, post-editing is often more cost-effective than a full human translation.
Need a human Dutch-to-English translation?
To discuss your text or project, get in touch via email or by filling in the contact form. For information about the cost of having your text post-edited, see the editing rates page.