Post-editing cost must be included

What it is

Total cost of machine-translation post-editing includes the reading, judgment, correction, validation, rework, and workflow upkeep required to bring MT output to a target quality, rather than comparing generation speed or per-word price alone. Machine-translation post-editing (MTPE) may raise productivity, but dense, subtle, or product-inappropriate errors can make it approach or exceed translation from scratch. Savings must be measured for a defined content class, language pair, quality level, and workflow.

Why it happens

An editor must decide whether fluent output is faithful before typing a correction, creating cognitive effort. Terminology, negation, reference, and UI-context errors may take only a few keystrokes to fix but much longer to investigate. Technical effort can be approximated with keystrokes or edit distance and temporal effort with processing time, while neither fully represents cognitive effort. Integration, data governance, terminology and prompt maintenance, quality sampling, fallback, and vendor switching add overhead. Savings are most likely when raw output fits the particular domain and language.

Studying it

• Use a crossover or matched design in which comparable translators handle difficulty-matched but non-repeated material under translation-from-scratch and post-editing conditions. Counterbalance conditions and order across participants and items to avoid carryover from seeing the same source segment twice.
• Record net processing time, pauses, keystrokes, edit distance, quality errors, and severity. Include system waiting and research time, not only active typing inside the editor.
• Compare at an equivalent target quality; faster output with more errors is not the same product. Also record editor experience, trust calibration, and fatigue.
• Report distributions and failure cases so average productivity does not hide negative returns for a content domain, language, or consequential string.

Where it stops holding

Small edit distance does not imply low cost or good quality: accepting a bad suggestion requires almost no typing but can cause serious harm. Organizations define “light” and “full” post-editing differently, so purpose, audience, and tolerated errors must be specified first. Translation-memory coverage, repetition, translator familiarity, and compensation model affect comparisons. Short pilots may miss terminology contamination, rework, training, and long-term maintenance.

Applying it

• Budget generation, post-editing, specialist review, in-context QA, project management, data compliance, tool maintenance, and rework separately; an MT invocation price is not total cost.
• Establish a baseline by language pair and content class and scale only when equivalent quality is achieved at lower total processing cost.
• Route poor output to translation from scratch or a higher review tier so editors are not forced to repair an unusable draft.
• Compare forecast and realized savings periodically and attribute terminology errors, withdrawals, support cases, and update work back to the relevant content batch.

Related

• Same group: S1.08.1 Machine translation suits high-volume, low-risk content · S1.08.2 UI microcopy has a higher error rate than long-form text · S1.08.3 Legal, safety, and medical copy must not rely on machine translation alone
• Adjacent: S1.05 Translation context and string reuse · L1.05 Human-in-the-loop intervention
• Search terms: machine translation post-editing · post-editing effort · ISO 18587