Template-Based Text-to-Image Alignment for Language Accessibility: A Study on Visualizing Simplified Text

This thesis provides a framework for generating accessible images from simplified text using structured prompt templates across multiple image styles. We tackle the need for visual illustrations that help users with cognitive disabilities and learning difficulties by combining simplified text with template-based image generation. Our strategy relies on four established text simplification datasets (OneStopEnglish, SimPA, Wikipedia, and ASSET) to produce a balanced corpus of 400 sentence pairs covering news, web, and encyclopedic domains. We developed five prompt templates—Basic Object Focus, Contextual Scene, Educational Layout, Multi-Level Detail, and Grid Layout—each with a different layout logic designed to adhere to cognitive accessibility principles based on established guidelines. Unlike previous methods that rely on short, fixed prompts, we introduce a systematic prompt engineering pipeline that transforms full simplified references into tailored accessibility-driven prompts. This approach maintains semantic alignment and captures real-world language variation. The method controls visual criteria such as object count, spatial distancing, boundaries, and content restrictions while modifying layout structures to improve understanding. We experimented with a wide range of image styles and assessed the five prompt templates for their influence on accessibility and overall clarity. To evaluate the effectiveness of the generated outcomes, we combined automatic metrics, including CLIP similarity scores, with detailed expert annotations across multiple evaluation dimensions. The final findings demonstrate how prompt structure, visual style, and content coverage influence the accessibility and overall quality of the generated images, marking a substantial contribution to cognitively accessible AI-generated content.