Documents and the editor

What it does#

This is where the actual work lives. Everything else, workflows, glossaries, memory, engines, exists to make what happens to a document better. A project holds one source language, one or more target languages, and belongs to a customer. Inside it, a document is one uploaded file, split into sentence-sized segments per target language. Localization OS reads 87 file formats, from Word, Excel, InDesign and FrameMaker to software resources, subtitles and Trados or memoQ bilingual files; the File formats page lists every one with what is translated and what is kept.

Each segment carries a status, untranslated, translated, needs review, or approved, and any QA flags a machine check has raised: a glossary or do-not-translate violation, damaged formatting tags, a number that changed, or a translation that reads as a refusal or leaked commentary rather than a real answer. The editor is the side-by-side review screen: source, target, translation-memory suggestions and QA findings together, so a reviewer can approve or correct in one place.

A document's status is derived from its segments in every language, never set by hand: new (nothing translated), in translation, in review (everything translated, not everything approved), and done (every segment approved). The translate buttons name the language they act on, and Translate all starts one background job per language.

Repeated segments follow the first approval. Approving a segment also applies its target to the document's other open repeats of the same source text in the same language. Each copy is checked against its own source for broken tags or placeholders first; a copy that would break is left alone and reported. A propagated copy shows a Propagated badge and is never approved automatically. Copies that are already approved, edited by a person, or locked from translation memory are never touched. Unlink repeats keeps one segment out, and the last propagation can be undone.

Formality, per language and down to one document. Set the register machine translation should use, per target language, as a customer default, on a project, or as an override on a single document; the most specific setting wins. Every language takes formal, informal or auto, and many take their own registers: Sie or du in German, vous or tu in French, usted or tu in Spanish, and casual, polite or honorific in Japanese, among others. The chosen register is one sentence in the prompt, recorded on the run, and also shown to the quality checks so a deliberate choice is not reported as an error. Nothing changes until someone sets a value.

Paused for review. While a workflow run is paused on a human review step, the editor shows the pending machine translation read-only under Waiting for review on run N, with a link to the run. Other writes to that document and language wait until the run resumes, so the reviewer and the pipeline never edit the same text at once.

Your file comes back as it went in. A file exported into a language with no translation is returned byte for byte. A translated file keeps its line endings, byte-order mark and, where the translation fits, its original encoding; see Encodings.

Translate comments, per project. Comments in Word, Excel (notes and threaded comments), PowerPoint, OpenDocument and Visio files are translated only when the project's Translate comments option is on. It is off for new projects; switch it on the project page (manager or admin), with project set-comments ID on, through the API or over MCP. It applies to files uploaded afterwards, and a comment's author and date never change. Projects created before the option existed keep translating Word comments until it is switched.

Right-to-left layout. For Arabic, Hebrew, Persian and other right-to-left targets, Excel and OpenDocument sheets are set right to left, InDesign and InCopy paragraphs get right-to-left direction and the World-Ready composer, and SVG text gets its direction and anchors mirrored. Left-to-right exports are unchanged.

Excel exports keep your workbook. Translated cells and text boxes are written into your own workbook package, so text boxes, shapes, pictures, charts and macros come back as uploaded. Workbooks uploaded before this change export their cells as before; upload again to translate their text boxes.

Plural forms. When a target language needs more plural forms than the file has, the missing forms are filled with the plural translation rather than left blank or in English. See QA checks.

Many files at once. Upload a ZIP and each supported file inside becomes its own document; files that cannot be used are listed with the reason. Export the batch as one ZIP with a folder per language and the original folder structure.

Getting started#

1. Upload the file into a project that has your target languages.
2. Apply 100% translation-memory matches first. It costs nothing and removes work.
3. Translate, or better, run a full workflow so the quality checks happen too.
4. Review whatever is flagged. The editor shows source, target, memory suggestions and the exact finding side by side.
5. Approve the segments you are happy with. Approving writes them to memory.
6. Export the file, or export the whole project as a delivery bundle.

doc upload 7 ./handbook.docx
doc apply-tm 12 de-DE
doc translate 12 --locale de-DE
doc export 12 --locale de-DE --out handbook.de-DE.docx

Works with#

• Workflows: a workflow always runs over one document and one target language, with a run view showing the path it took, step by step.
• Translation memory: approving a segment writes it to memory; applying 100% matches reads back from it in bulk, for free.
• Glossaries: glossary and do-not-translate violations are found on a document's segments and surface as QA flags in the editor.
• Memory: correcting a translation here and saving is the evidence memory later distills into rules.
• Evaluations: a document can be snapshotted into an evaluation dataset, a frozen copy including its current translations as references.
• Engines: translating a document dispatches per-segment engine calls; a translation-memory hit skips the engine entirely.
• QA checks: every machine translation is checked for tags, placeholders, numbers, terms, ambiguous dates, line breaks and joined words.

Current limits#