FormA11y — Research and Development of a Tool for Remediating PDF Forms for Accessibility
Sparsh Paliwal, Joshua Hoeflich, J. Bern Jordan, Rajiv Jain, Vlad I. Morariu, Alexa Siu, Jonathan Lazar · 2025 · ACM Transactions on Computer-Human Interaction, Vol. 32, No. 1 · doi:10.1145/3702317
Summary
Paliwal and colleagues — a team spanning the University of Maryland's Maryland Initiative for Digital Accessibility (MIDA), Northwestern, and Adobe Research — tackle a problem that has been largely overlooked in PDF accessibility research: the remediation of interactive *PDF forms*, not just long-form reading documents. Although roughly 2.5 trillion PDFs exist online and as few as 2.4% are accessible, prior tooling work (e.g., PAVE, A11y by Pradhan et al.) has focused on headings, reading order, and table tagging for research papers, leaving the dense, field-heavy world of government and healthcare forms unaddressed. The authors begin with interviews with two expert PDF remediators (7 and 10 years' experience) and audits of accessible government forms to identify three core pain points: the remediation process is non-intuitive (tooltip rules differ for textboxes, checkboxes, radios, and table fields), repetitive (forms may have 100+ fields), and overwhelming because of high information density. They then design FormA11y, a React-based web tool that uses existing ML models (Form2Seq, LayerDoc, FUDGE) to generate a partially remediated baseline, and then walks the user through three distinct, color-coded steps — Fields, Groups, and Tooltips — each showing only the elements relevant to that step. Labels are drawn and associated via OCR text selection rather than being typed manually. The authors evaluated FormA11y against Adobe Acrobat Pro in a within-subjects study with 20 participants (4 expert, 16 novice) using two 80-90 field VA forms from the Flamingo dataset seeded with controlled synthetic errors across fields, groups, and tooltips.
Key findings
FormA11y users remediated forms in 41.0% of the time needed for Acrobat (mean 12:55 vs 31:29, p=6.54×10⁻⁹); 100% of FormA11y participants finished within the 40-minute cap versus only 70% for Acrobat. Overall task performance — remaining accessibility errors after remediation — improved by 82.4% with FormA11y. Broken down: Field errors dropped from mean 2.5 to 0.15 (p=0.0002); Group errors from 4.6 to 0.3 (p=0.0003); Tooltip errors from 14.1 to 3.3 (p=0.0002). Precision and recall rose across all three steps (Fields: 100% / 98.3% precision/recall vs 92% / 79.4% for Acrobat). Crucially, participants introduced dramatically fewer *new* errors with FormA11y (0 new field errors vs 0.7, 0 new group errors vs 2.5, 2 new tooltip errors vs 7.2). System Usability Scale scores averaged 83.4 (SD 7.7) for FormA11y — well into the 'better, more usable products' range — versus 45.6 (SD 19.5) for Acrobat, which fell below the 70 passable threshold. Qualitative findings attribute the improvements to color-coded field types (orange textboxes, pink signatures), drag-to-select label association replacing manual typing, visible group boundaries (red outlines), and step-based chunking reducing cognitive load. The group step remained the hardest in both tools. Acrobat still bested FormA11y on one niche capability: editing arbitrary custom tooltip text, which FormA11y can only append.
Relevance
For accessibility practitioners in government, healthcare, legal aid, and education — sectors where PDF forms remain the dominant workflow for applications, permissions, intake, and benefits — this paper is directly actionable. It quantifies how much of the accessibility-debt problem is actually a *tooling* problem: the same professionals using Acrobat produce 5-10× more residual errors than they do with an interface designed around the remediation task. The Department of Justice's 2024 Title II rule now explicitly requires WCAG 2.1 AA conformance for state and local government PDFs, so expected demand for form remediation is rising fast. The paper also reframes the human role: rather than generating accessibility metadata from scratch, the remediator becomes a *reviewer* of ML output, which aligns with the trajectory of Adobe Sensei, Form2Seq, and FUDGE models (currently 10-34% error rates). Limitations: the study used only two VA-style forms with controlled synthetic errors and university-recruited participants rather than enterprise remediators; FormA11y does not yet validate reading order, links, alt text, or sensory characteristics; and the tool's treatment of sub-field character inputs (as in Figure 22) remains unsupported. The ML backbone was mocked rather than deployed, so real-world accuracy will depend on model quality.
Tags: PDF accessibility · PDF forms · PDF/UA · document accessibility · remediation tools · born-accessible · blind users · screen readers · machine learning · usability study · accessibility tooling
Standards referenced: PDF/UA · WCAG · Section 508 · Matterhorn Protocol · WCAG 2.1 Level AA · ADA Title II