Mesa Translation

Technical translation and localisation, explained.

Machine Translation Post-Editing: What It Costs and When It Pays

Machine Translation Post-Editing: What It Costs and When It Pays

Machine translation post-editing — running content through an engine and paying a linguist to correct the output — is now the default offer from most vendors. It is genuinely cheaper for some content and quietly more expensive for others, and the dividing line is not where buyers usually assume.

The economics only work above a quality floor

Post-editing saves money when the raw output is good enough that correcting it is faster than translating from scratch. Below that threshold the editor spends longer untangling a confident, fluent, wrong sentence than they would have spent writing a correct one. Engine quality varies enormously by language pair and domain, so "we use MT post-editing" is not a price; it is a question you have to answer per language.

Light and full post-editing are different products

Light post-editing fixes errors that change meaning and leaves everything else alone. The result is accurate and often slightly odd. Full post-editing brings the text to the standard you would expect from human translation. Vendors quote both, buyers frequently compare a light price against a full expectation, and the disappointment is predictable.

Where it reliably pays

High-volume, repetitive, low-risk content is the sweet spot: support articles, product listings, internal knowledge bases, release notes. These share a useful property — the cost of an awkward sentence is low, and the volume makes even modest savings material.

Where it reliably does not

Marketing copy, user interface strings and anything legally binding. UI strings are short and context-free, which is exactly the condition under which engines guess worst. Marketing depends on tone, which is the first thing post-editing flattens. Contracts need a human who can be held responsible for the reading.

Measure edit distance, not satisfaction

If you run post-editing at any scale, ask for edit distance — how much of the engine output survived. It turns a subjective argument about quality into a number you can track by language pair, and it tells you when an engine has drifted before your editors start complaining.

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *