Multi-Column Reading Order
Which block comes next — the sequencing decision that turns detected columns into readable text.
Multi-column reading order is the sequencing decision within columnar layouts: given the page's detected blocks — columns, headings, figures, captions, sidebars — in what order does a reader (human or machine) traverse them? It is the ordering half of multi-column parsing (detection being the other), and the half where subtle judgment lives: column boundaries can be geometrically obvious while the sequence remains genuinely ambiguous — does the full-width heading start a new flow across both columns beneath it, or cap the flow above? Does the boxed sidebar interrupt the reading or accompany it?
The general reading-order entry covers the task's full scope; the multi-column case has its own recurring structures worth naming. Standard columnar flow: down column one, top of column two — with the sequence direction script-dependent (right-to-left scripts reverse the column order). Interrupting elements: full-width titles, figures, and tables that segment the columns into vertical zones, each zone flowing independently — the model that misses the zoning threads text through a figure's position. Non-flow elements: sidebars, pull-quotes, marginalia, and advertisements that belong to the page but not the sequence — best emitted as attached-but-separate content, since forcing them into the linear order corrupts it wherever they land. Continuation structure: the newspaper's jump lines and the form's "see reverse," where the flow leaves the page entirely and resumes elsewhere — document-level ordering that per-page models can't see.
The evaluation and consumption points mirror the parent entries: sequence metrics against annotated orderings, linguistic-coherence checks as cheap production verification, and the consumers' stake — chunking, extraction context, text-to-speech for accessibility, and every LLM that will read the serialization — inheriting whatever order was chosen. In columnar documents more than anywhere, reading order is the parse: the same detected blocks in the wrong sequence are a different, wrong document.
Read down, not across — parsing the layouts where naive left-to-right produces word salad.
The sequence a human eye would follow — inferred from layout, and essential before a word is serialized.
Before reading the words, a model has to see the page the way a human does — headings here, table there, footnote at the bottom.
Proof Perimeter runs document AI inside your own perimeter — with a provenance record on every field.
Book a demo