Reconstruct sequence before translation
A full-page map plus labeled crops for columns, cards, captions, and notes.
Region relationship
Document layout sets sequence. Translation cannot repair a false source order.
Sequence check
Read the first and last line of every region before joining the text.
Correct words can still form a wrong result
Reconstruct the source reading order before translation, using labeled crops as the audit trail.
Map the visual regions
Mark the page's title, columns, boxes, captions, sidebars, footnotes, and repeated headers. Decide the order a human reader would follow. Do not assume top-to-bottom coordinates describe every document.
Mixed-direction text and rotated labels may need separate regions. Chat bubbles should be ordered by time and speaker, not by left and right position.
Extract region by region
Create one crop per logical region and retain small overlaps where sequence could be disputed. Name the crops in reading order. Remove repeated headers only after confirming they are identical and not part of the content.
Paste source text into a structured draft with headings that identify each region. Check sentence boundaries and list indentation before asking for translation.
When the source uses cells and headers, rebuild their relationships with the workflow to translate a table from an image.
Validate the repaired sequence
Read the first and last line of every region against the full image. Confirm that captions stay with their images, footnotes with their markers, and field values with their labels.
Translate only after the source sequence makes sense without invention. If layout remains ambiguous, deliver separate translated regions rather than one artificially smooth paragraph.
Repair scrambled OCR reading order from columns, cards, captions, sidebars, chat bubbles, tables, and forms before translating the text. Continue with the troubleshooting guide collection for adjacent workflows.