Combining Semantic Clustering and Sentence Embedding Representation for Abstractive Biomedical Summarization

Automatic biomedical text summarization is an important task aimed at providing essential information from biomedical text. This helps medical professionals in accessing the most relevant significant points in remarkably short amounts of time. Unlike extractive summarization, which selects relevant sentences from the source document, abstractive summarization involves generating the summary word-by-word, which not necessary present in the original document. Applying abstractive summarization to specialized domains like biomedical text remains a challenging task due to the sensitive content, complex jargon of these fields, and contextual window limitations of large language models in the case of long documents. Abstractive summarization also suffers from the redundancy of multi-sentence summaries. To address these challenges, we propose a novel four-step method that combines semantic clustering and sentence embedding representation able to generate the abstractive summary. These steps are: (i) preprocessing and sentence embedding representation, (ii) clustering, (iii) summarizing the obtained clusters, and (iv) aggregating the sub-summaries. Through these steps, our proposed method is able to capture semantic similarity within the text, resulting in a coherent and informative summary. Additionally, our method allows to preserve the context within the same cluster. To show the effectiveness of our method, we perform an automatic evaluation over PubMed Summarization Dataset using the Recall- Oriented Understudy for Gisting Evaluation (ROUGE) metrics. The obtained results demonstrate that the proposed summarization system outperforms various baselines and existing state-of-the-art systems.