Machine Translation Post-Editing (MTPE)

Machine translation has improved markedly in recent years, but improving has meant its errors changing character rather than becoming fewer. Raw output now reads fluently, which makes its mistakes the kind that are easier to miss: a sentence invented out of nothing, a negative turned into a positive, a term chosen wrongly for the context. Catching precisely these insidious errors is what MTPE is worth.

Which texts are suitable, and which are not?

Not every text suits MTPE. Texts with a high rate of repetition, plain phrasing and settled terminology come out of a machine well: product descriptions, the routine sections of user manuals, internal documentation. In creative or ambiguous texts, or texts with legal consequences, machine output is both inadequate and risky. A contract or an advertising strapline calls for a specialist translation, not MTPE.

That is why the first step is always an assessment of suitability. If a text is not suited to MTPE, saying so at the outset prevents problems later. Machine-derived translation is likewise not recommended for documents intended for official use; for those, the officially approved translation processes set out in our document guide apply.

How the process works

The MTPE process does not end where the machine finishes; it begins there. The raw output is a draft, and the real labour is the human scrutiny that makes that draft dependable. At the centre of the workflow, therefore, is a reading of the output against the source. Throughout, the aim is to take the machine’s speed while weeding out the insidious errors it produces. The steps below follow the same order whether the level chosen is light or full; only the degree of rigour at each step changes. The suitability assessment right at the start is decisive, because putting the wrong text into MTPE renders every subsequent step pointless.

1. Suitability assessment. The type and purpose of the text are examined, and it is decided whether MTPE is appropriate at all.
2. Choice of level. Light or full post-editing is decided according to where the text is going.
3. Terminology preparation. The client’s glossary and preferred terms, where they exist, are set up; a machine does not know these on its own.
4. Post-editing. The specialist reads the output against the source, dealing first with invented passages, omissions and terminology errors.
5. Final check. Figures, names and negations are verified separately; these are where machines go wrong most often.

Delivery and points to watch

Depending on the level chosen, the text delivered is either dependable but plain, or natural and ready to publish. Setting the expectation correctly at the start matters: it is unreasonable to expect a text ordered as light post-editing to read as fluently as a human translation. What each level promises is therefore stated explicitly in the agreement.

The most critical issue specific to this service is confidentiality and data security. A text pasted into a free online translation tool has, in effect, been sent to that service’s servers. For documents containing confidential information or personal data, that is unacceptable. Where the engine used in MTPE runs, and whether the data is retained, are concrete questions that should be asked; for work requiring confidentiality, open-source solutions that do not leak data are preferred.

Human translation or MTPE?

Choosing between MTPE and a specialist translation is usually a matter of balancing speed and budget against quality. For a high-volume, short-lived text meant for internal use, MTPE is sensible; for a text that will be the face of the brand, stay published for a long time or carry legal consequences, translating directly is safer. The decision is made text by text; one document from an organisation may suit MTPE while another does not.

The language pair affects the choice as well. Machine translation performs noticeably well in some pairs and weakly in others; the suffix structure and word order of Turkish make the output more fragile in certain pairs. An MTPE approach that works in one language pair may therefore fall short in another. The target language deserves as much weight as the type of text when the decision is made; where MTPE is not appropriate, the honest recommendation is direct translation rather than the promise of a saving.

What determines the price?

Whether MTPE is economical depends on the quality of the machine output. If the output is good, little editing is needed and the cost falls; if it is poor, editing turns into something close to retranslation and the expected saving disappears. MTPE is therefore not always cheap; it can only be said to be cheap once the text and the language pair turn out to be suitable. An honest quote is not given without seeing the output.

To find out whether your text is suited to MTPE, you can send us a sample section through the quote form. You will find the price bands on the pricing page, and the direct translation options for texts that are not suitable on the translation services page.

Situations that come up often

• The wrong kind of text. MTPE requested for advertising or legal material; machine-derived output is not recommended for these.
• Confidential data. A document containing personal data having already been run through an open online tool; where the source goes should be questioned at the outset.
• Mismatched expectations. Full post-editing quality expected at the light post-editing price; the level and what it promises must be clear from the start.
• Hidden errors. Fluent-looking output that turns a negative into a positive; this is why figures, names and negations are checked separately.

The information here is general; current requirements should be confirmed before you apply. We recommend checking the present position with the authority handling your case.