Using Generative AI to Create Text- and Audio-Based Summaries of Educational Textbooks

This bachelor thesis investigates the use of generative artificial intelligence for automated processing of educational textbooks by first generating textual summaries and then using these summaries to produce audio-based learning content. The motivation for this work stems from the difficulty many school students face when engaging with traditional educational textbooks in Ukraine, which often present dense, static text that is not adapted to diverse learning needs, as well as from the high effort required to manually create alternative explanations and supplementary materials. The project proposes an automated pipeline that processes school textbook chapters by extracting their structure, segmenting the content into chapters, and generating rewritten explanatory narratives for each chapter using large language models. Rather than performing simple summarization, the generated texts aim to preserve the essential factual content of the original material while presenting it in a clearer, more engaging, and student-friendly form. The generated explanations are additionally converted into audio format using textto-speech technology to support multimodal consumption of the content. A controlled experimental methodology is employed to explore how different prompting strategies and generation parameters influence the quality of the generated summary content. The evaluation focuses on semantic similarity between the original textbook content and the generated explanations, using established automatic metric. The experimental results show average BERTScore F1 values between 0.71 and 0.74 across all evaluated conditions; based on these findings, it can be concluded that prompt design constitutes the primary factor influencing semantic similarity between the original textbook content and the generated explanations, while the effect of temperature variation within the tested range remains limited.