Machine learning based lay summarization in biomedical domain

Communicating scientific research to the general public poses both a vital and challenging endeavor. Bridging the gap between scientific knowledge and public understanding often necessitates the use of lay summaries, which offer simplified versions of research findings. In this thesis, we explore the effective communication of biomedical research to the general public by generating concise lay summaries. Drawing upon curated datasets from prominent biomedical journals, namely PLOS and eLife, we develop a system that generates lay summaries from the abstract and main text of research articles. This thesis explores two distinct approaches for Biomedical Lay Summarization. In Model 1, the input articles undergo text preprocessing techniques including the removal of non-ASCII characters and redundant parentheses. The Stanza library is employed to identify and tag bio-entities, utilizing named entity recognition models specifically designed for biomedical text, such as BC5CDR and JNLPBA. Model 2, on the other hand, utilizes the original abstract text as a summary and combines it with summaries generated by a BART-based model trained on preprocessed full-text articles. To accommodate the model's input length limit of 1024 characters, longer articles are divided into sections and fed sequentially into the model. The model's output is then concatenated to form the complete summary. Post-processing techniques, including text cleaning and the use of the Stanza library for unknown word identification and tagging, are applied to refine the generated summaries. These two approaches contribute to the advancement of automatic summarization systems in the biomedical domain, facilitating the comprehension and dissemination of scientific knowledge to a broader audience.