Unsupervised Paper2Slides Generation

Abstract

Although presentations are an excellent medium for sharing academic opinions and ideas, there has been a scarcity of research into automating the "paper to slides" generation task, and a lack of publicly available datasets. In response, we propose an inventive optimization framework based on reconstruction loss, harnessing cutting-edge Large Language Models (LLMs) and unsupervised learning. This approach facilitates the creation of high-quality slide decks from scientific papers, offering heightened adaptability and flexibility. Our evaluation results provide empirical evidence of our model’s superior performance in comparison to baseline models.