Ontology-Driven Transformations for PDF Form Accessibility
Utku Uckun, Ali Selman Aydin, Vikas Ashok, IV Ramakrishnan · 2020 · Proceedings of the 22nd International ACM SIGACCESS Conference on Computers and Accessibility (ASSETS 2020) · doi:10.1145/3373625.3418047
Summary
This demonstration paper presents TransPAc (Transformations for PDF Form Accessibility), an assistive technology that transforms PDF forms into accessible HTML pages so blind screen-reader users can fill them out. PDF forms present multiple accessibility barriers: many are scanned images (non-interactive), so blind users can read text but cannot enter data into form fields. Even interactive PDF forms have poor screen reader support — the serial reading order does not match the visual layout, making it difficult to associate labels with their correct form fields. Additionally, Adobe Acrobat automatically shrinks font sizes to fit entered data within fixed field dimensions, making the filled-out form unreadable to sighted colleagues who later review it. TransPAc addresses all of these issues through a five-stage pipeline: (1) Element Extraction — detecting text and form field elements using both Adobe Acrobat Pro and Amazon AWS Textract in combination for better accuracy, annotating each with 2D bounding box coordinates; (2) Form Labelling — analyzing geometrically close text elements to generate a suggestion list of possible labels for each form field; (3) Logical Segmentation — using an ontology-based rule system to group horizontally aligned elements into logical lines and vertically aligned lines into logical segments, preserving the spatial relationships of the original PDF; (4) PDF to HTML Transformation — topologically sorting the logical segments and converting them to a linear HTML document with proper form fields and label suggestion lists; (5) Populating PDF Form Fields — transparently transferring data from the HTML form back into the PDF, with an optimization algorithm that selects the largest readable font size and expands field dimensions if needed.
Key findings
The system leverages the fact that blind users are already proficient with web-based screen reader interaction, including HTML form navigation shortcuts, which are far better supported than PDF form accessibility. The ontology encodes computational rules about horizontal and vertical alignments, closeness thresholds, and visual cue detection (such as black boxes separating form sections) to drive the segmentation — ensuring the screen reader narrates all elements in a logical group before moving to the next, which is essential for understanding form context. The suggestion list for form field labels was shown in prior work (a user study with 14 blind participants across 856 PDF form documents) to help users understand what data a field expects. The font optimization algorithm for populating the PDF ensures that filled-in data is not truncated or shrunk to unreadable sizes — if data cannot fit, the algorithm greedily expands the field by grabbing adjacent white space. The resulting filled PDF maintains visual fidelity for sighted users who receive the completed form.
Relevance
TransPAc addresses one of the most persistent and practical accessibility barriers in computing: inaccessible PDF forms are ubiquitous in government, healthcare, education, finance, and employment. Blind users frequently encounter PDF forms for tax filings, medical intake, college applications, and workplace onboarding — and current screen reader support for PDF forms is fundamentally inadequate. The approach of transforming PDF content to an accessible HTML intermediate representation, rather than trying to fix PDF accessibility directly, is pragmatic and leverages existing screen reader capabilities. For accessibility practitioners, the paper highlights that even "accessible" interactive PDF forms often fail because of label-field association problems that screen readers cannot resolve. The ontology-driven segmentation approach could potentially be applied to other document transformation challenges beyond forms. Limitations include the short demo format (3 pages), reliance on the accuracy of element extraction tools, and the assumption that the PDF has machine-readable text (scanned image-only PDFs would need OCR first).
Tags: screen readers · document accessibility · PDF accessibility · blindness and low vision · web accessibility · assistive technology · software accessibility