ISO 5060: Evaluating Translation Output

Output evaluation is a separate question

A project can follow a defined workflow and still produce a defective sentence. Conversely, observing one good sentence says little about the rest of a delivery. Evaluation needs to say what was examined, what counted as a problem and what the observations justify.

This is the distinction between evidence about a process and evidence about a result. Both may matter, but they should not be substituted for each other.

An illustrative assessment

Imagine a translated installation guide with three possible findings: a missing prohibition, an inconsistent product term and an awkward but understandable phrase. Our suggested exercise is to record the source passage, target passage, reason for the finding and proposed correction for each.

Before assigning severity, state what the guide is for and how an error could affect the reader. Then have another qualified evaluator review a small shared sample. A disagreement about classification can reveal an unclear instruction or a missing project requirement.

This is an original educational example. It does not reproduce an ISO scorecard or prescribe a universal threshold.

Sampling and the meaning of a result

Our practical recommendation is to make the limits of a sample visible. A review of selected passages should not be described as a complete inspection. Record the selection method, size and exclusions, including any parts that received a different type of check.

For comparisons, keep the evaluation setup sufficiently aligned. A score from a short, deliberately difficult sample should not casually be compared with one from a random selection of routine content.

Connection to LISA, MQM and AI

The relationship to LISA is a shared concern with explicit quality assessment. MQM provides a framework for selecting and classifying error types; this ISO reference is guidance about evaluation. They should not be treated as identical publications.

Because the published scope focuses on human evaluation, an automated score alone should not be labelled “ISO 5060 certified.” AI can support an editorially designed assessment, but claims about conformity require more than naming the reference.