ISO 18587: Human Post-editing of Machine Translation

Why post-editing needs its own description

Reading a generated translation and repairing it is a distinct production activity. It involves deciding whether the target preserves the intended meaning, whether omissions or unsupported additions exist and whether the result meets the assignment. The source and project requirements remain important even when the first target draft arrives instantly.

Those are practical considerations, not a reconstruction of the standard's detailed requirements. A complete implementation must use the applicable full text.

An illustrative workflow problem

A help article says, “Do not reset the device while the update is running.” An automated translation accidentally removes the prohibition. A post-editor who reads only the target might find a fluent sentence and miss the reversal.

For this scenario, a sensible editorial exercise is to compare source and target instructions, identify the changed action and record the correction. The team should then ask whether similar source constructions appear elsewhere. A repaired sentence is useful evidence, but it does not establish that the whole project has been checked.

What the draft changes in our reporting

The new draft title is relevant to discussion of AI and other automated output. It does not mean a final replacement edition has been published. We report the draft separately and do not assign it the requirements or authority of a completed standard.

Connection to LISA and AI

The continuity is an interest in defined production and accountable quality. It is not a claim that ISO 18587 descends directly from LISA QA. For AI-assisted work, our recommendation is to record who checks what, which resources they receive and how acceptance is decided.

An automatically revised translation is not, merely by being revised, evidence of full human post-editing. A quality score also cannot establish whether the intended human process took place.