Towards More Accessible Scientific PDFs for People with Visual Impairments: Step-by-Step PDF Remediation to Improve Tag Accuracy

Research Background and Issues

• What problems or challenges did the authors identify?
  Only a small fraction of PDF files meet the accessibility standards for screen readers used by blind individuals (PDF/UA). These accessibility issues are particularly severe, posing significant barriers to learning and career development in STEM fields. Additionally, existing PDF remediation tools are often difficult to use, especially for non-experts, as their workflows are complex and prone to errors.

• Why is this issue important?
  As the primary format for scientific literature, PDFs play a crucial role in academic communication. Failure to improve PDF accessibility effectively will further exacerbate the underrepresentation of visually impaired individuals in scientific research and career advancement. This issue also violates multiple laws and regulations (e.g., Section 508 and the European Accessibility Act), which mandate that public documents must be accessible to everyone.

• Research Motivation and Related Work
  Previous studies and existing literature show that less than 2.4% of scientific PDF files fully comply with accessibility standards. Tools like Adobe Acrobat Pro are complex and difficult to use, with limited quality in automatically generated tags. Furthermore, research has confirmed that users find the accessibility remediation of PDFs with mathematical formulas and complex layouts particularly challenging. Therefore, there is an urgent need to develop a user-friendly tool with high-accuracy tagging capabilities.

Solution

• What methods or solutions did the authors propose?
  The authors developed PAVE 2.0, a semi-automated PDF remediation tool offering eight clear steps to improve tag structure and enhance accessibility. It includes AI-supported features, such as a generator for alternative text for mathematical formulas, and an interactive user interface designed to minimize the need for direct manipulation of the PDF structure tree.

• What are the innovative aspects of this solution?

  1. Introduced an AI-driven math editor to assist in generating alternative text for mathematical formulas based on MathSpeak rules.
  2. Proposed a "step-by-step remediation process," dividing tasks into content-independent and content-dependent steps, enabling users to complete remediation progressively from simple to complex tasks.
  3. Designed a completely new accessibility scoring standard, including 13 manually checkable criteria, to reliably evaluate the accuracy of PDF tags.

• What are the implementation steps and key technologies used?

  1. Define regions: Use AI models to detect areas such as paragraphs, headings, formulas, and graphics, allowing manual adjustments.
  2. Configure reading order: Modify the order using linear drawing tools or lists.
  3. Set heading levels: Automatically or manually adjust the hierarchy of headings.
  4. Create tables and lists: Define subcells by drawing dividing lines.
  5. Assign alternative text to images: Support word count optimization for alternative text.
  6. Provide a math editor: Generate LaTeX formulas using AI models and convert them into alternative text.
  7. Review pages and set metadata: Offer a preview of the PDF with tag structures included.

Research Outcomes

• What specific outcomes were achieved?

  1. PAVE 2.0 significantly improved participants' tagging accuracy, increasing it from 42.0% to 80.1% for experienced users and from 39.2% to 75.2% for beginners.
  2. The study found that 15 out of 19 participants preferred using PAVE 2.0 in the future, and all participants recommended it as suitable for beginners.

• What advantages does it have compared to existing solutions?

  1. More intuitive user interface: Compared to Adobe Acrobat's default complex structure tree operations, it greatly reduces the barrier to use.
  2. More systematic and efficient remediation process: Tasks are divided into steps, guiding users to complete them in a logical sequence.
  3. AI-assisted functionality for generating alternative text for mathematical formulas, which is absent in Adobe Acrobat Pro.

• What are the experimental or evaluation results?

  1. Users preferred to keep the remediation time per page within 2-5 minutes, and PAVE 2.0 largely met this requirement.
  2. When using the automatic tagging feature, PAVE 2.0's initial tag quality (56.9%) exceeded Adobe Acrobat Pro by 19.4 percentage points.
  3. PAVE 2.0 was able to correct errors in automated model-generated tags, including heading structures, mathematical formulas, and tables.

• Limitations and Future Directions

  1. The current version does not yet support complex structures, such as multi-page tables and multi-column lists.
  2. The list drawing and reading order configuration features require further optimization, particularly in identifying AI tagging errors and training users to use these tools.
  3. Future research should focus on applying this process in diverse publishing environments and exploring long-term user feedback in real-world scenarios.
  4. Integrate MathML standards and develop interactive solutions for accessible mathematical formulas, such as tree-structured navigation.

Conclusion

This study makes significant contributions to the field of PDF remediation, not only optimizing the accessibility process but also proposing new solutions for handling complex content such as mathematical formulas. PAVE 2.0, with its intuitive interface and AI support, significantly improves tag quality and proves to be particularly user-friendly for beginners and general users. This work has the potential to redefine the standards for PDF accessibility remediation, especially in the academic communication domain.