Why a PDF manual is hard to read by machine

A Portable Document Format (PDF) file describes how a page looks, not what its parts mean. A manual's headings, warnings, tables, and steps reach software as text and graphics placed on pages, unless the creator added a separate layer of structure tags. Where a manual cannot leave a secure room, that structure has to be rebuilt on the local machine, and a wrong rebuild can leave a warning detached from the step it governs.

A PDF stores glyphs, not structure

The Library of Congress describes PDF text as encoded mainly by identifying glyphs within fonts, not by Unicode code points. When the mapping to Unicode is missing, extracted text can be unreadable.

Document structure is a separate layer, and it is optional. A PDF's logical structure is stored apart from its visible content, with pointers between the two, and it exists only if the creator or the creation process adds tagging. Files made by printing to PDF generally carry no tags. Without that layer, the page content does not say that a line is a heading or that a block of text is a table cell.

The format is specified in ISO 32000, a standard of the International Organization for Standardization (ISO). PDF 2.0, ISO 32000-2, was published in July 2017 as the successor to ISO 32000-1:2008 (PDF 1.7), and a dated revision followed in December 2020, according to the Library of Congress format description.

Tags add structure only when someone adds them

Tagged PDF was introduced with PDF 1.4. It adds a structure tree whose elements use standard types such as paragraph, heading, list item, figure, and table data cell, and the order of the tags defines the reading order. Because the structure is stored separately, the order of logical elements can be independent of where graphics objects sit on the page.

PDF/UA-1 (PDF Enhancement for Accessibility) is an accessibility profile of PDF 1.7, published as ISO 14289-1 in 2012 with a minor update in December 2014. As the Library of Congress summarizes it, all real content must be tagged in a logical reading order, layout and pagination objects are marked as artifacts and left out of the structure tree, figures need alternative text, and headings, tables, and lists have detailed usage rules.

Tags do not appear on their own. A 2004 article cited in the Library of Congress entry on the PDF family states that automated tagging alone is unlikely to give satisfactory results unless a document is very simple. A tool that depends on tags needs a plan for files that have none.

Without tags, software infers the structure

When tags are missing, software infers structure from how the page looks. The DocLayNet paper calls layout analysis a key requirement for high-quality PDF conversion. The dataset it introduces holds 80,863 manually annotated pages labeled with 11 classes: Caption, Footnote, Formula, List-item, Page-footer, Page-header, Picture, Section-header, Table, Text, and Title. Manuals are one of its six document categories.

The authors argue that earlier datasets drawn only from scientific article repositories lack layout variety, and that models trained on them lose accuracy on more varied layouts. The detection models they tested still scored roughly 10% below the level at which human annotators agree with one another.

Three limits follow from the label set itself:

• No class exists for a warning or caution. The labels name kinds of page region, such as Text or Picture, not roles in a procedure.
• A box labeled Table marks a region. Recovering rows, columns, and header cells is a further task. A survey of table work describes pairing object detection, which finds the table, with table structure recognition methods to build a working system.
• The annotation guidelines pair each Caption with one Picture or Table. A sentence elsewhere that refers to a figure by number, or a numbered callout inside an illustration, has no label of its own.

Reading order is modeled as a task too. The LayoutReader paper treats reading order detection as foundational for understanding visually rich documents. The authors built ReadingBank, a set of 500,000 document images with reading order, text, and layout information, taking the order from Word document XML metadata. They report that their model significantly improves both open-source and commercial optical character recognition (OCR) engines at ordering text lines.

What this means for a technical manual

Consider a manual page where a warning box sits above the step it governs, a table's header row defines its cells, and a numbered callout ties an illustration to a parts list. In a tagged file, those relations can be recorded in the structure tree, if the tags are present and correct. In an untagged file, software has to infer them from geometry, using models trained on page collections that may differ from the manual in hand.

A reader that reports the reading order it chose, and the page location of everything it extracts, lets a person check the result against the page.

What to ask

• Are the manuals tagged PDFs, and were the tags checked against the page?
• For untagged files, which layout model infers the headings, tables, and reading order, and was it tested on manual pages?
• Does the tool report the reading order it chose?
• Does every extracted item carry its page location, so a person can check it against the page?
• Does a warning stay attached to the step it governs in the extracted text?

Azcuna's TRACE is one offline viewer for S1000D, MIL-STD-3001 XML, and PDF technical manuals. It keeps the source location and the markings of everything it shows and never modifies the source. It was proposed to NAVAIR under Navy SBIR Phase I topic DON26BZ05-NV078, and the proposal was submitted on September 23, 2026.

Sources

1. 1PDF (Portable Document Format) Family · Library of Congress
2. 2PDF 2.0, ISO 32000-2 (2017, 2020) · Library of Congress
3. 3PDF/UA-1, PDF Enhancement for Accessibility, Use of ISO 32000-1 · Library of Congress
4. 4DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis · arXiv (Pfitzmann et al., KDD 2022)
5. 5LayoutReader: Pre-training of Text and Layout for Reading Order Detection · arXiv (Wang et al.)
6. 6Deep learning for table detection and structure recognition: A survey · arXiv (Kasem et al.)