iTagPDF: Towards Finally Automating PDF Accessibility

Paper Title

iTagPDF: Towards Finally Automating PDF Accessibility

Publication Info

• Topic area: Automating accessibility metadata generation for research paper PDFs.
• Keywords: PDF accessibility, semantic tagging, LaTeX, reading order, metadata, vision-language models, object detection, OCR, accessibility remediation, document processing.

Background and Problem

• Problem / challenge: PDF accessibility remains a persistent issue due to the lack of automated tools for generating accessibility metadata. Existing methods rely heavily on manual tagging, which is tedious, error-prone, and inconsistent.
• Significance: Ensuring PDF accessibility is critical for enabling assistive technologies to interpret content for people with disabilities, yet most academic PDFs fail to meet accessibility standards.
• Motivation and related work: Prior tools like Adobe Acrobat and open-source alternatives require manual intervention and often produce low-quality tags. Efforts to create borne-accessible PDFs, especially in LaTeX, have been limited in scope. Vision-language models and object detection pipelines have potential but fail to fully address the semantic and structural tagging needs of PDFs.

Solution

• Proposed approach: iTagPDF, a system that combines visual rendering and source document semantics to automate PDF accessibility tagging.
• Novelty:
  1. Combines visual and source representations to generate accurate tags, reading order, and metadata.
  2. Introduces a new dataset and metrics tailored for evaluating PDF accessibility.
  3. Establishes a strong baseline for automated tagging performance.
• Procedure and key techniques:
  1. Process PDF pages using object detection and OCR to identify visual content regions.
  2. Parse LaTeX source documents to extract semantic context.
  3. Align visual and semantic data using heuristics and a large language model (LLM).
  4. Generate and embed 13 document-level tags (e.g., <H1>, <P>, <Figure>) and 3 content-specific tags (e.g., <TH>, <TD>) into the PDF.
  5. Refine reading order and generate metadata for tables, figures, lists, headings, and formulas.

Results

• Concrete findings:
  • iTagPDF achieved high accuracy in bounding box localization (95.28%), classification (95.96%), and reading order (97.26%).
  • Outperformed Adobe Acrobat and author-submitted PDFs in accessibility tagging, with an average normalized score of 95.19%.
  • Demonstrated stable performance across multiple runs, with a normalized accuracy of 94.06%.
• Advantage over baselines:
  • iTagPDF surpassed Adobe Acrobat and authors’ manual tagging in reading order, headings, lists, tables, and captions.
  • Generated more consistent and accurate metadata, avoiding common errors like redundant tags and misclassified artifacts.
• Experiments / evaluation:
  • Evaluated on a dataset of 40 PDFs (554 pages) from ACM CHI and ASSETS conferences.
  • Introduced new error metrics (Box Error Rate, Classification Error Rate, Reading Order Error Rate) for manual evaluation.
  • Compared against Adobe Acrobat’s auto-tagging and author-submitted PDFs.
• Limitations and future work:
  • Focused on LaTeX-authored PDFs; needs extension to other formats like Microsoft Word.
  • Relies on heuristics and LLMs, which may fail in ambiguous cases (e.g., figures, captions).
  • Does not yet support certain tags (e.g., <Link>) or inline formulas.
  • Future work includes creating larger training datasets, improving object detection models, and reducing pipeline latency.

Summary

iTagPDF is a novel system that automates PDF accessibility tagging by combining visual and semantic representations. It outperforms existing tools and author submissions in generating accurate tags, reading order, and metadata for research paper PDFs. Evaluated on a dataset of ACM conference papers, iTagPDF demonstrated high accuracy and consistency, marking a significant step toward fully automating PDF accessibility remediation. Future work aims to extend its applicability to other document formats and improve its robustness and efficiency.