Human Post-Editing for AI Book Translation: What to Review First

Human post-editing should begin with errors that can invalidate the translation, not with sentences that merely sound awkward. Confirm that the complete target book matches the frozen source, then review omissions, additions, reversed meaning, names, terminology, numbers, and cross-references. Only after those checks should an editor spend time on voice, rhythm, and polish.

That order matters because fluent AI output can still contain a missing paragraph or a confident mistranslation. A perfect sentence cannot repair an absent chapter.

Use this sequence:

1. Define what the translated book will be used for.
2. Freeze the exact source and target files.
3. Reconcile structure and completeness.
4. Correct meaning-critical and controlled information.
5. Review terminology and continuity across the whole book.
6. Edit voice and target-language naturalness in connected passages.
7. Test the exported file and recheck every major fix.

Download the AI book post-editing worksheet to record the decision, evidence, severity, owner, and recheck status for each issue.

Choose the required finish before you edit

“Post-editing” does not identify one fixed amount of work. A private reading copy and a commercial edition have different acceptance thresholds.

ISO 18587:2017 covers full human post-editing of machine-translation output and post-editor competence. It is not a label for any quick cleanup. If your project needs only a readable draft, call that a limited review and state exactly what was not checked.

This is also where the distinction in AI versus human book translation becomes operational: the method matters less than who owns the unresolved risks.

Freeze the evidence before changing the target

Keep three separate files:

• the exact source used for translation;
• the untouched AI output; and
• the working post-edited version.

Record the source filename, revision, language, target language, translation date, output format, glossary version, and any workflow settings that changed the result. If the source changes during review, create a new source revision and assess the change. Do not silently mix paragraphs from two editions.

Before editing prose, compare:

• opening and closing content;
• chapter and section count;
• table of contents order;
• headings, lists, tables, captions, notes, appendices, and back matter;
• first and last paragraph of every major section; and
• source-language text that may be intentionally retained or accidentally untranslated.

The broader translation QA checklist for books and long documents provides a release-level issue log. The post-editing worksheet in this article adds the decision about whether to keep, repair, or retranslate AI output.

Review in an order that changes the release decision

Use four passes. Each pass answers a different question.

Pass 1: Is the output complete and paired with the right source?

Look for missing, duplicated, added, truncated, or reordered material. Verify chapter boundaries before line editing. Check that captions remain with their figures and footnotes still point to the intended notes.

These are blockers because later editing can make an incomplete target look finished.

Pass 2: Does the target preserve the source meaning?

Prioritize passages where a small error changes an action or claim:

• negation: do versus do not;
• modality: must, may, should, and can;
• direction: increase versus decrease;
• sequence: before versus after;
• scope: all, some, only, and except;
• numbers, dates, units, percentages, ranges, and identifiers; and
• attribution: who said, did, believed, or caused something.

For example, if the source says a character almost confessed, an AI target that says the character confessed is a meaning error even if the sentence reads naturally. If a maintenance instruction says to disconnect power before opening a panel, changing the sequence is a blocker, not a style issue.

Pass 3: Are controlled decisions consistent?

Search the whole target for:

• character and place names;
• titles and forms of address;
• product and organization names;
• defined technical terms;
• abbreviations and their expansions;
• invented words;
• quotations and cited titles; and
• do-not-translate items.

Record the approved form and rejected variants. Do not correct one occurrence and assume the rest are safe. Use the book translation terminology template to turn a one-time edit into a reusable decision.

Pass 4: Does it read as the intended target-language book?

Now review paragraphs and scenes without staring at the source after every sentence. Check register, rhythm, paragraph cohesion, pronoun reference, dialogue differentiation, repetition, and genre conventions. Then return to the source and confirm that stylistic edits did not change meaning.

This two-view method prevents opposite failures: source-shaped translationese when the editor never reads the target independently, and elegant mistranslation when the editor stops comparing with the source.

Use severity instead of polishing everything equally

Classify each issue by consequence.

The MQM error typology separates accuracy, terminology, linguistic conventions, style, locale conventions, audience appropriateness, and design/markup. Use only the categories that change your project decisions. A detailed taxonomy is useful; a score that lets one blocker disappear inside hundreds of clean segments is not.

Decide whether to keep editing or retranslate

Post-editing is not automatically cheaper than starting again. Run a representative sample before committing to the whole book:

1. Select the opening, a dialogue-heavy passage, a terminology-dense passage, a table or note, and a difficult late chapter.
2. Record minutes spent reading, researching, editing, and rechecking.
3. Count blockers and repeated major patterns.
4. Estimate whether the same repair will recur across the book.

Retranslate a section when the editor must reconstruct meaning sentence by sentence, pronouns or subjects cannot be traced reliably, terminology is systematically wrong, or the output structure no longer maps to the source. Switch the entire workflow when representative samples show that repair cost is dominated by retranslation rather than correction.

Do not publish a universal threshold such as “edit if fewer than 20% of words change.” Word-change percentages do not measure whether the changed word reversed a warning or merely improved punctuation.

Separate correction from approval

The person making edits can miss the pattern that produced them. Add a final pass that starts from the candidate release, not the editor's working memory.

• Recheck every blocker and major issue against the source.
• Search the target for all rejected terminology variants.
• Compare the final table of contents and structure again.
• Open the actual EPUB, PDF, or DOCX in the application readers will use.
• Review metadata, links, notes, figures, and navigation.
• Record who approved language, subject matter, and production.

Full-document context also matters in evaluation. A large professional-annotation study of machine translation gave reviewers full document context for MQM assessment; isolated sentence review can hide reference and continuity problems.

Where BookTranslator fits

BookTranslator can create a structured translated draft from a complete EPUB, PDF, or DOCX, apply an automatic glossary, and provide bilingual output where supported. That can reduce the work of moving chapters through short text boxes and make important passages easier to compare.

It does not replace a qualified editor or domain reviewer. Upload the clean source through the book translation workflow, keep the untouched output, and use the worksheet to decide where human judgment is required. For scan-heavy sources, fix extraction problems before treating the target as a post-editing candidate.

Final post-editing gate

Do not approve the book until all of these are true:

• the exact source and untouched AI output are retained;
• complete-file reconciliation found no unresolved blocker;
• meaning-critical passages, numbers, names, terms, and references were checked;
• target-language editing covered connected passages, not only isolated sentences;
• every major correction was searched for as a possible recurring pattern;
• the final file was tested in its delivery format; and
• the intended use, review depth, reviewer ownership, and known limitations are recorded.

Human post-editing is valuable when it concentrates judgment where the AI draft is weakest. It fails when “human reviewed” is used as a vague label for an undefined amount of cleanup.