Abstract
Transforming complex biomedical texts into accessible lay summaries is a critical endeavor in Natural Language Generation (NLG). This study addresses the challenges associated with this task by employing a multi-aspect approach. Firstly, we undertake a comprehensive analysis of discourse structures within a diverse range of biomedical datasets and clarify underlying patterns and structures. Secondly, we designed the power of prompting strategies to integrate training on these varied datasets, thereby reducing the noise introduced by their diversity. This twofold strategy fine-tunes the model’s training and enriches it with the ability to generate coherent and simplified lay summaries of biomedical content. Our experimental results clearly demonstrate the effectiveness of our study, underscoring its potential to make complex medical information more accessible to general readers.