Translate
The Translate step is in Preview: ready to use, and it may still change based on feedback.
Category: Processing
The Translate step translates each page's text into one language and keeps the translation beside the page. Reviewers who do not read the document's language can switch the page to the translation in the Transaction Viewer, and fields marked Translate are extracted in that language, with the text on the page kept beside each value (Extraction Fields).
Place it after OCR and before Extract: it translates the words OCR read, so the translation is tied to where each phrase is on the page. A workflow has at most one Translate step.
Settings
| Setting | What it does | Default |
|---|---|---|
| Target Language | The language pages are translated into. Translated fields are extracted in it too. | English |
| Model | The model for the translate task. Empty uses your organization's default, then the system default. | Empty |
| Skip Pages Already in the Target Language | A page the OCR engine reports in the target language, with no text in another script, is not sent to the model. | On |
| Send Page Images | Sends the page image with its text to a model that accepts images, which helps with tables and stamps. | On |
How a page is translated
- The page's lines are split at column gaps, so a translation never runs into the next column. The model groups the words into phrases (a table cell, a label, a line of an address) and translates each one. Numbers, codes and text already in the target language are left as they are; an amount printed against its label ("Total:1,250") stays in the label's translation, and an amount written out in words is translated.
- On a mixed page, only the text not in the target language is translated. A page that needs no translation is recorded as such.
- One page is sent per call, with the text of the pages before and after it as context.
- A page the model cannot translate usefully is stored untranslated, with a warning in the transaction's Activity Log; the step goes on with the other pages.
Cost
The step's model usage is recorded and shown per step in the transaction's cost, like Extract and Classify. Translate this page in the viewer and the live translation of an edited field are recorded under the same Translate row.
Redaction
A Redact step after Translate also removes redacted values from the translation: a phrase that lost a word loses its translation too.
Without a Translate step
Fields marked Translate are still extracted, translated by the model into English on its own, and the workflow shows a warning. In the viewer, a person with edit rights can translate a page on demand with Translate this page (in the image panel's ⋮ menu).