Abstractive Biomedical Long Document Summarization Through Zero-Shot Prompting

Abstractive summarization involves generating a coherent and concise summary that captures the main topics of the original document. In this work, we address the task of abstractive long single-document summarization in the biomedical domain, which faces two main challenges. Firstly, biomedical documents are often very long and contain a significant amount of information. Secondly, there is a lack of annotated datasets for biomedical text summarization tasks. In response to these challenges, we introduce a novel unsupervised method that follows an extractive-abstractive summarization framework. In the first stage, extractive summarization is modeled as a graph where nodes represent sentences and edges represent the semantic similarity between sentences. This graph is optimized to select salient sentences that highlight the main topics of the document. In the second stage, we leverage the potential of GPT-3 through zero-shot prompting to generate an abstract from the extracted content. Experimental results on the PubMed dataset, evaluated using the ROUGE metric, demonstrate the effectiveness of our approach, achieving competitive performance compared to recent supervised deep learning-based state-of-the-art methods.