Ideal Answer Generation for Biomedical Questions using Abstractive Summarization
Iqra Naveed, Muhammad Wasim · 2023
Finding precise information from biomedical literature is challenging because of the information overload and the ever-increasing size of the biomedical literature. Biomedical experts struggle to find precise information by reading complete documents, so there is a need to develop query-based summarized answers from the relevant articles. One solution to overcome this challenge is the development of question-answering (QA) systems to provide biomedical experts with precise information in the form of a summary (also referred to as ideal answer). Although extensive studies exist to find the answers to biomedical questions as facts (also known as factoids), the work on ideal answer generation is limited. In this study, we introduce a methodology for generating rephrased summary answers for biomedical questions from the relevant articles using the benchmark BioASQ dataset. We compare three transformer-based models, namely: BigBird, BART Large CNN, and Long T5 pre-trained models of abstractive summarization for generating ideal answers from biomedical snippets. We evaluate these models using the well-known ROUGE metric. Our experiments suggest that the BART Large CNN model outperforms other transformer-based models, achieving an average score of 0.428, 0.304, and 0.376 for ROUGE-1, ROUGE-2, and ROUGE-L respectively.