WTPDF / PDF/UA-2 Examples by the LaTeX Project

[davidcarlisle]

WTPDF / PDF/UA-2 Examples by the LaTeX Project

The following files demonstrate various aspects of Well Tagged PDF documents conforming to PDF/UA-2.

They were all generated with LuaLaTeX (lualatex-dev in TeX Live 2024).

The files are a mixture of small examples demonstrating specific features, older out of copyright documents that have been re-typeset as tagged pdf, and contemporary documents including recently published arXiv papers, course notes, and conference papers.

The files here are all PDF 2.0. PDF 1.7 versions of the same documents are available from PDF/UA-1 Examples by the LaTeX Project.

---

Access to the Files

The full collection of PDF files is available at Google Drive, where you may select one or more individual files to download, or, at the
top of the page is a Download all link which will generate a zip file and download the full collection.

Google drive directory of all example PDF files

The LaTeX sources are available from this repository also, where appropriate, we link to the original files used as source material.

---

Verification of PDF/UA-2 compliance

There are not yet many validators that correctly handle UA-2 (given that the standard was released in March 2024 not that
surprisingly). One online validator you can try on the smaller examples is

VeraPDF — PDF/A and PDF/UA Validation

Please note that some PDF viewers modify the PDF when opening it (to allow for annotations, for example). In some cases this is known to break the PDF/UA-2 standard. If that happens re-download and use a different viewer.

---

The Samples

Simple Examples with MathML Associated files

All three conform to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

Three small examples demonstrating the use of Associated Files to Tag mathematics. Each formula is associated with two associated files. A LaTeX fragment representing the original source, and a MathML document.

mathml-AF-ex1

mathml-AF-ex2

Sample-AF-Math-LaTeX

amsmath LaTeX package documentation

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

The amsmath package defines the main markup structures for mathematics in LaTeX.

This manual has examples of many kinds of aligned equations and similar structures. This version has been enhanced to produce Well Tagged PDF.

amsldoc-tagged

tagpdf LaTeX package documentation

Conforms to: PDF/UA-2 PDF/A-4 WTPDF/Accessibility WTPDF/Reuse Arlington

The tagpdf LaTeX package is a core part of the LaTeX support for tagged PDF.

Its documentation already conforms to WTPDF and PDF/UA-2 and a snapshot is included here.

tagpdf

ArXiv publications

Tagged using MathML extracted from the arXiv-supplied html versions of the documents.

They were each submitted to arXiv under a CC Licence permitting re-use such as this experiment, The tagged documents are available under the same licence.

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

2401.09965v1-tagged — Original Source

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

2401.09436v1-tagged — Original Source

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

2401.05361v1-tagged — Original Source

Niels Bohr: The Theory of Spectra and Atomic Constitution; Three Essays

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

These essays by Niels Bohr are available as LaTeX source from The Project Gutenberg.

Additional TeX markup has been added to produce Tagged PDF. Also all math expressions were converted to MathML using LaTeXML.

47464-t-tagged — Original Source

William Shakespeare: MACBETH

Conforms to: PDF/UA-2 PDF/A-4 WTPDF/Accessibility WTPDF/Reuse Arlington

macbeth-tagged — Original Source

This document uses a provided LaTeX source of the play text. The LaTeX markup has been enhanced to produce Well Tagged PDF.

American Standard Version of the Bible (1901 text)

Conforms to: PDF/UA-2 PDF/A-4 WTPDF/Accessibility WTPDF/Reuse Arlington

The plain text source of the ASV Bible, 1901 as provided by Wikisource. This has been marked up as LaTeX to generate well tagged PDF. This example demonstrates a custom role map with structured tagging corresponding to the Testament/Book/Chapter/Verse structure shown in this work.

ASV Bible — Original Source

DEIMS 2024 Conference paper

Conforms to: PDF/UA-2 PDF/A-4 WTPDF/Accessibility WTPDF/Reuse Arlington

The paper Enhancing LATEX to Automatically Produce Tagged and Accessible PDF submitted to DEIMS 2024, Tokyo.

As well as describing the approach to PDF tagging used for these examples, the paper does itself form an example of tagging a
contemporary conference paper. This is the version as prepared for the TeX Users Group publication, TUGBoat.

tb139mitt-deims24

The presentation at the DEIMS conference including a demonstration is available as a video.

PDF Association sample poster

Conforms to: PDF/UA-2 PDF/A-4 WTPDF/Accessibility WTPDF/Reuse Arlington

An article describing the PDF Association work on accessibility produced for the PDF Association launch of Well Tagged PDF.

pdfa-art

Sample Chemistry/Math notes

This is a small contemporary document used as notes on mathematical aspects of Chemistry.

In this example, the math is associated with just LaTeX source Associated files, not MathML.

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

525Da-23-group-theory

A small template exam paper.

Conforms to: PDF/UA-2 PDF/A-4F WTPDF/Accessibility WTPDF/Reuse Arlington

PHY-exam

Wilhelm Busch: Max and Moritz

Conforms to: PDF/UA-2 PDF/A-4 WTPDF/Accessibility WTPDF/Reuse Arlington

A LaTeX document that does not have math and the main language is not English. Showing tagging of images, verse structures and the use of more than one (marked up) language in a document.

pg17161-tagged — Original Source

[petervwyatt]

Awesome work! But I did note 2 files with errors and a few other issues:

• for 2401.09965v1-tagged.pdf

  • There are 2 x Ref objects (objects 2682 and 2705) in the structure tree which are not structure element dictionaries but file specification dictionaries (for associated files). This is incorrect. Table 355 requires Ref to be structure elements. This is an error.

• for PHY-exam.pdf

  • Page 3, Widget annot for the button is missing AP (appearance stream info), as required in PDF 2.0. See Table 166 in ISO 32000-2:2020.

• for 2401.09436v1-tagged.pdf

  • has deprecated ProcSets a few times (not technically an issue as future PDF/A-4 dated revision will permit deprecated features, but could save some space).
  • Has Type 1 FontDescriptor/CharSet a few times - also deprecated in PDF 2.0 (not technically an issue as future PDF/A-4 dated revision will permit deprecated features, but could save some space).
  • in Adobe Acrobat, all the embedded Mathml XML files show as "Size = 0.00 bytes" - if you set the Size entry in the F/UF Params dictionary then I think the correct size will display in Acrobat's file list nav pane. You might want to consider adding if that works...

Just for discussion: several files have private PTEX entries for XObjects, etc. such as PTEX.FileName and PTEX.InfoDict which can include author, filename, etc as per the pdfTEX documentation (https://texdoc.org/serve/pdftex-a.pdf/0). Since PDF/A files are intended for long-term preservation, this has the potential to cause issues for FOIA and similar requests since the presence of private data might slip past various redaction workflows. A modern equivalent is to use an XMP Metadata stream instead of 2nd class custom PDF keys which makes this more discoverable.

[u-fischer]

@petervwyatt Thanks for the report.

I will add an appearance to PHY-exam.pdf.

But while testing that I got two curious complains from arlington (I used the lastest verapdf version) from the attached pdf:

• It complained that a DA key is missing. Why? There is no variable text, the text is inside the appearance.
• It complained that the Ff bitmask is wrong for a radio field. Why that? This is a pushbutton so naturally it has not a Radio bitmask.

test-utf8.pdf

[FrankMittelbach]

@davidcarlisle one question: does LaTeX not warn for 2401.09965v1-tagged.pdf that it needs one more run or was this simply overlooked?

[davidcarlisle]

@FrankMittelbach yes it did warn but got lost in all the tagging debug logging, so user error, me :( If you see the build script in the sources it now grep's the log and re-runs as needed after an update this morning so that shouldn't happen in future.

[petervwyatt]

@u-fischer I can't speak to veraPDF's Arlington implementation (I'm not 100% sure which version they used and whether they augmented the rules, or how they determine what each object is) but DA messages are most likely related to PDF Errata #323 as this is not well specified.
When I ran against the latest Arlington PDF Model (in GitHub, using my PoC C++ test harness) I only got the messages I listed above (+ other messages that I checked and confirmed as noise/limitations of my implementation) - I did not get any DA or Ff messages..

[davidcarlisle]

PHY-exam.pdf has been updated

[bdoubrov]

@u-fischer I can't speak to veraPDF's Arlington implementation (I'm not 100% sure which version they used and whether they augmented the rules, or how they determine what each object is) but DA messages are most likely related to PDF Errata #323 as this is not well specified. When I ran against the latest Arlington PDF Model (in GitHub, using my PoC C++ test harness) I only got the messages I listed above (+ other messages that I checked and confirmed as noise/limitations of my implementation) - I did not get any DA or Ff messages..

@u-fischer @petervwyatt

• Ff bitmask error is indeed a bug in veraPDF implementation. It is fixed in the latest dev build 1.25.278
• DA is still required for push buttons according to Arlington model: FieldBitPush.tsv. This might indeed change after PDF Errata #323 is resolved. But for the moment, as far as I understand, Arlington model follows Table 228 of ISO 32000, where DA is specified as required. Though I'm not sure if Push buttons fit into the category of fields that contain variable text.

[ErroneousBosch]

Speaking as an accessibility professional ( am no expert in LaTeX), the dependence on VeraPDF is not wise. While it seems to be able to verify that tags exist in a nominal structure, the quality and usefulness of the actual tags and structure being generated sub-standard. The lack of ActualText in math equations means this fails to meet the PDF/UA-2 or WTPDF standards. Tables are very baseline and primitive, without any header cells or scoping. Image captions are not contained correctly and alt text seems to have some issue where it is not being picked up by screen readers.

It is premature to claim any level of real compliance. All of these issues are ones that there is no automatic checker for and can only be picked up through human testing.

[ErroneousBosch]

I'm in the process of gathering the information, and need time to injest ISO 32000-2:2020. I work at an academic institution, and we are approaching this from a policy/legal compliance standpoint, in our case WCAG 2.1 AA and Section 508. More importantly, we have to test for demonstrable accessibility which both the example files above and files we generated with TeXLive 2024 do not meet.

Screen reader performance was especially poor, checking with Apple VO, NVDA, and Adobe's own reader. That is honestly where the rubber meets the road. Compliance to a standard that isn't implemented anywhere isn't a useful compliance, especially if it means not meeting real-world accessibility needs.

Like I said, I am gathering more useful details to submit in one or more issues.

[josephwright]

Compliance to a standard that isn't implemented anywhere isn't a useful compliance, especially if it means not meeting real-world accessibility needs.

True to some extent, but one issue in this area is that without good examples (complex inputs meeting the agreed standards in terms of structure), viewers, etc., will not be developed that can read them. So simply saying 'target only what is readable now' doesn't really work: current reader implementations have significant gaps in coverage

[davidcarlisle]

Note that this collection is specifically a collection of examples for PDF/UA-2, that is the new PDF 2.0 based standard. It is explicitly here to allow implementors of PDF consuming software to have a collection of documents to test against. So while, to produce an accessible document for end users today, you do need to target PDF 1.x and PDF/UA-1 this collection is to allow us to test PDF/UA-2 generation and allow consuming applications to have a set of PDF/UA-2 documents to test.

[u-fischer]

"real-world accessibility" for eg math is currently not given anyway: in 1.7 not as the standard has no support for it, in 2.0 not as the implementations do not support it. Our files are meant to push development forward. I just was at the PDF week where we used the files to demonstrate and discuss the implementation problems.

[car222222]

@ErroneousBosch wrote:

we are approaching this from a policy/legal compliance standpoint,
in our case WCAG 2.1 AA . . .

This is relevant, since ensuring (as far as possible) that the PDF output is compliant
with the latest WCAG level AA provisions is definitely important.
But (a very big BUT:-) there are some severe limitations on what can be done
to achieve such compliance within the LaTeX processing, since:

1. PDF is not a “web-technology”, so that much of WCAG needs reinterpretation.
   Also, it may not be possible to achieve much “within PDF itself”, since PDF is
   “only a format”, and in many areas the PDF specification mandates very little
   concerning how processors and associated AT behave.
   Thus, very much unlike WCAG, the PDF standard does not prescribe how, and
   even whether, any processor (or associated AT) concretely implements anything:
   for example, most of its very detailed provisions concerning how to produce visual output are described only in terms of various models, and not in terms of actual
   physical actions, or specific code to be interpreted outside these abstract models.

2. It is not feasible to apply many WCAG provisions to “the PDF format and PDF producers alone”, since much of it largely concerns the capabilities and behaviour of consumer applications (i.e., how they interpret the format in order to present a PDF document to users); this includes, of course, their use of AT.

Therefore, for much that is important to compliance with WCAG provisions, it is impossible for PDF production software alone to ensure, or even provide support
for, these requirements.

We probably need to look at this, to see if it helps to improve our WCAG compliance:
favicon.ico

and Section 508.

And it would be unwise for us to go anywhere near supporting such purely local "legal requirements”!

[FrankMittelbach]

Am 14.05.24 um 01:54 schrieb ErroneousBosch:

More importantly, we have to test for demonstrable accessibility which both the example files above and files we generated with TeXLive 2024 do not meet.

The big question here is why are they not meeting it? Because there are errors in them with respect to implementing UA-2 or because consuming software up to now is not capable of properly handling PDF 2 structures yet? So far industry hasn't bothered much with the improved structures provided by PDF 2 (and necessary for higher quality accessibility) because there were (nearly) no documents that used them --- so so why bother if there is no use case?

Screen reader performance was especially poor, checking with Apple VO, NVDA, and Adobe's own reader. That is honestly where the rubber meets the road. Compliance to a standard that isn't implemented anywhere isn't a useful compliance, especially if it means not meeting real-world accessibility needs.

all true, but if your road is currently a gravel surface with huge holes in it, the question is: do you want to continue running over it only with noisy tanks because everything else that would be a comfortable car will break down, or do you strife for improving the road? right now accessibility of PDFs is so poor because consuming software is based on 1.7 and UA-1 + a lot of heuristics (which differ from implementation to implementation and therefore also do not give a good user experience over all). Now, by producing UA-2 docs you to not get magically better accessibility, in fact you are likely to get even worse, because consumer software handles the improved structures badly or not at all and their heuristics fail with such documents. But as it was pointed out, the goal of producing documents that comply to the new (and better) standards, was to make showcases of where in the current consumer software fails with PDF/UA-2 and this way drive good implementations of the new standard in the consumer software. With the ability of providing a corpus of complex documents that meet PDF/UA-2, we are fairly confident that this could happen and in fact we already see movements in this respect

Like I said, I am gathering more useful details to submit in one or more issues.

please do, but also please keep in mind the purpose of the generated documents, e.g., - things in which we go wrong should be improved on our end to make the documents better - but things that go wrong in consumer apps because they do not understand the standard, should really (with some pressure) communicated by the community to the vendors.

[Jeffrey-Kuan]

I'm at the workshop right now! Can I request feedback for this PDF I made? The LaTeX source code uses amsthm, which has many votes to be made compatible:

https://tailorswiftbot.godaddysites.com/arxiv-2209-11114

[ErroneousBosch]

@Jeffrey-Kuan Interesting results. Some of the alt text is better, and it is at least there.

Looking at page 3:

• The math formula themselves are Formula tags, but are themselves embedded text, so the text is being read instead of the alt-text of the formula tag. This means the math is not coming through clearly. Usually this is better solved by use of the "Actual Text" attribute on a P tag contained in the Formula tag, but you'd need to test things.
• Looking at the alt text that is on the Formula tags: q to the plus or minus 1 times the quantity: q to the 1 minus 2 n; plus q to the 2 n minus 1 is better than expected, is the alt text being manually entered?
• the one above it, mathcal L super 1 2 plus mathal L super 2 3 plus dot dot dot mathcal L super L minus 1 L is bringing through a LaTeX tag into the alt text.

The matrix representation's alt text is not usable, possibly cut off:

a 16 by 16 matrix whose entries are almost all 0. Each diagonal entry is a star. There are 20 nonzero entries which will be listed in blocks. The 2-5 entry and 5-2 entry both have a term...

My reader skipped the element entirely.

Tagging structure is a mixed bag, since body text is separate from contained formulas. The real killer is that the reading order is completely useless. Inline formulas are sometimes read after their containing text, some text is not completely read, references are read inline as numbers, and footnotes behave very erratically.

Overall still a hard fail.

[u-fischer]

@Jeffrey-Kuan There are a number of errors.

• You are using somewhere undeclared tags, this shows up as lines as this in the PDF /UNKNOWN<</MCID 6>> BDC and should also gives warnings from tagpdf.

• Not all chars have ToUnicode values. verapdf and PAC complain about this.

• There are lots of text parts which aren't tagged at all (they are not surrounded by BDC/EMC markers in the PDF), for example the "and" between the two formula in the abstract. This indicates error in the coding.

• You have quite often P-tags nested inside P-tags. This is not allowed and you should get warning from a current tagpdf.

• The bookmarks of the PDF looks odd as you didn't exclude the tagging commands.