Correct words can still produce the wrong narrative order
Judge word recognition and reading order separately. A page can contain correctly recognized words in a sequence that changes its meaning.
The working result is a region map that explains what should be read first, what belongs together, and which captions, sidebars, headers, or footnotes require separate handling.
Put the procedure into practice with PDF to Text. Continue with OCR, Accessibility and Reading Order and OCR Output Formats: TXT, JSON, TSV, hOCR, ALTO and PDF.
Continue with the structured OCR and page-layout collection for tables, forms, coordinates, and output formats.
Reading-order inspection
- Mark every page region
- Trace the main reading path
- Label sidebars and captions
- Handle headers and footers consistently
- Test copy order or assistive reading order
Map regions before joining their text
Map regions before reading lines
Mark the intended start and end of each column, then identify content that interrupts the main flow. Do not infer order from the plain-text export after headers and sidebars have already been mixed into paragraphs.
Use continuity tests
Read the last sentence fragment in one block and the first in the proposed next block. Grammar alone is not enough; headings, numbered steps, citations, and page continuation also provide sequence evidence.
Separate furniture from content
Running headers, footers, folios, crop marks, and repeated navigation can be retained as metadata without being inserted into every paragraph. Record the rule so the same material is treated consistently.
Test accessible sequence
For a searchable or remediated PDF, inspect the programmatic sequence rather than the visual page. W3C guidance emphasizes that meaningful order must be determinable when order affects meaning.
Build a region map before joining text
Take a page with two columns, a pull quote, a photograph caption, and footnotes. Draw region boundaries and write the intended reading sequence before looking at OCR output. Decide whether the caption belongs after the image reference and whether footnotes follow each column or the complete body.
Recognize regions separately when a full-page pass interleaves lines. Keep page coordinates and region identifiers during recombination. Then read the flattened output from beginning to end without the page image. Correct words in the wrong sequence still communicate the wrong document, so order needs its own acceptance check.
- Map regions first
- Keep captions and notes attached to context
- Test the exported sequence independently
Region handling
| Region | Common error | Review decision |
|---|---|---|
| Two body columns | Lines alternate across columns | Finish one column before the next |
| Sidebar | Inserted mid-sentence | Place at a meaningful boundary or label separately |
| Caption | Detached from its figure | Keep an explicit figure association |
| Running header | Repeated inside body text | Store as page metadata or omit by rule |
There may be more than one valid sequence
Magazines, forms, sidebars, captions, and footnotes can support task-dependent reading orders. The visual page and intended audience must guide the final sequence.