Post-editing and review of AI-generated translations

A translation can look perfect and still be wrong

Today's machine translation and artificial intelligence systems can produce very natural-sounding text. Fluency, however, does not guarantee that the content corresponds exactly to the source. Some of the most important errors are precisely the hardest to detect: grammatically correct sentences, apparently appropriate terminology and convincing prose that, when compared with the source, reveals a change in meaning, an omission or an interpretation that the original does not support.

Professional post-editing and review are intended to identify these problems and ensure that the final text is faithful, consistent and appropriate for its intended purpose.

Post-editing is more than language correction

Correcting grammar and improving the style of an AI-produced translation is not enough to determine whether the translation is correct. The reviewer must return to the source text.

A bilingual review makes it possible to check whether the content has been conveyed in full, whether logical relationships have been preserved, whether the terminology is appropriate, whether there are omissions or additions, and whether apparently equivalent words actually express the same concept in that particular context.

The risk of apparent fluency

Current systems are particularly good at producing plausible language. That very ability can make certain errors difficult to recognise.

A translation may be perfectly idiomatic and still turn a possibility into an obligation, remove a condition, resolve an ambiguity that the source leaves open, or choose a technically plausible but incorrect term.

In legal translation, for example, lexical correspondence does not guarantee conceptual correspondence. A term may have an obvious literal translation and still belong to a legal concept that has no direct equivalent in another legal system.

Particular importance in legal, business and technical texts

The greater the consequence of an error, the greater the level of control required. Contracts, pleadings and other procedural documents, corporate documentation, financial reports, business communications and technical documentation often contain specialised terminology and relationships of meaning that cannot be assessed in isolation.

This logic is consistent with the European Commission Directorate-General for Translation's approach to quality, which is risk-based: texts in which an error may have legal or procedural consequences require more extensive controls than lower-risk content. The Commission also takes account of the text's purpose, context and target audience.

What professional review involves

The level of intervention always depends on the quality of the material received. In some cases, machine translation provides a suitable basis for post-editing. In others, certain passages require substantial rewriting or fresh translation from the source.

The aim is not to preserve as much of the machine-produced text as possible. It is to obtain a correct final text. This means systematically checking correspondence with the source, specialised terminology, internal consistency, references, numbers and dates, register, syntax and suitability for the required language variety.

An internationally recognised professional activity

Post-editing is not simply a later read-through of machine-translated text. ISO 18587:2017 sets requirements for the process of full human post-editing of machine translation output and for the competences of the professionals who perform it. ISO itself indicates that this edition is to be replaced by a revision currently under development, covering post-editing of non-human translation output.

AI as a tool, not a seal of quality

The use of artificial intelligence and professional translation are not incompatible. The European Commission Directorate-General for Translation uses machine translation technology and allows its translators to use eTranslation output as an optional resource.

The official eTranslation page itself distinguishes between raw machine translation and human-quality translation, stating that where maximum accuracy is required, the output should be reviewed by a professional translator.

The decisive question, therefore, is not simply whether AI was involved. It is how it was used, for what kind of document and what level of control was applied to the result.

Confidentiality and data protection

There is another issue to consider before entering documents into external artificial intelligence systems: what those documents contain. Contracts, court proceedings, business documentation or administrative records may contain personal data or confidential information.

Whenever the use of a tool involves processing personal data subject to the General Data Protection Regulation (GDPR), principles including lawfulness, purpose limitation, data minimisation and accuracy under Article 5 must be considered. The tool used, its processing terms and the sensitivity of the content should therefore be assessed before documents are sent to an external service.

Technology in the service of translation

Artificial intelligence can speed up tasks, facilitate research and provide a useful tool in the translation process. But rapidly producing a plausible sentence and determining whether that sentence is the correct translation are different problems.

Professional review focuses on the second problem: checking meaning, context and terminology and applying critical judgement to the final text.