Topic: Investigation and Application of LLMs and Vision Models to Enhance Document Accessibility

Context: With the introduction of the European Accessibility Act, public institutions face the challenge of making all digital content accessible. This includes ensuring legacy documents (PDFs, Word files) are compatible with assistive technologies. Currently, a significant portion of university documents does not meet these standards. This thesis explores how modern AI technologies can bridge this gap.

Project Vision: The goal of this thesis is to investigate the potential of Large Language Models (LLMs) and Vision Models to automate or assist in the creation of barrier-free documents. Unlike rigid rule-based systems, Generative AI offers new possibilities for understanding context, describing images, and interpreting document structures.

Scope & Potential Directions: Students are encouraged to bring their own ideas to the table. The specific focus of the work can be tailored to your interests, but potential research avenues include:

·Automated Remediation: Developing a prototype that utilizes the university’s internal uniGPT model to automatically suggest structural tags or repairs for legacy documents.

·Visual Context: Using Vision Models to generate context-aware "Alt-Text" for complex diagrams and images.

·Evaluation & Verification: Investigating if LLMs can act as "auditors" to reliably detect accessibility violations that standard tools miss.

·Human-in-the-Loop: Designing a workflow where AI assists staff members in remediating documents, rather than fully automating the process.

Your Tasks:

·Literature review on the state of the art in AI-driven accessibility.

·Conceptualization of a solution architecture using available resources (e.g., uniGPT).

·Implementation of a Proof of Concept (PoC).

·Evaluation of the solution regarding accuracy, efficiency, and compliance with accessibility standards (e.g., WCAG).

We Offer:

·Creative Freedom: The opportunity to shape the direction of the thesis and define your own implementation strategy.

·Access to Infrastructure: Use of university-hosted models (uniGPT) and computing resources.