Scientific text summarization by attention-augmented deep clustering and semantic-aware sentence rating

One of the biggest challenges facing researchers is the rising amount of scientific literature. Automatic summarization offers an effective solution by generating concise summaries that aid comprehension and accelerate analysis. For lengthy documents, extractive summarization is a common method; however, selecting sentences that accurately convey the document’s meaning remains challenging. This study proposes an improved approach that leverages context comprehension to rank sentences. To better capture complex semantic structures in scientific text, deep clustering is enhanced using an Attention-Augmented Deep Embedding Clustering (ATT-DEC) mechanism built upon a Scientific BERT (SciBERT) encoding backbone, with sentence-level relevance scores assigned by a Multi-Layer Perceptron (MLP). Furthermore, a non-linear mathematical model is used to balance selection. Recall-Oriented Understudy for Gisting Evaluation (ROUGE) F1 metrics were used to assess the method on the ArXiv and PubMed datasets. On ArXiv (R1 = 0.4337, R2 = 0.2402, RL = 0.3527) and PubMed (R1 = 0.4508, R2 = 0.2585, RL = 0.3724), the results demonstrate improvements over state-of-the-art baselines across most evaluation conditions, highlighting the model’s potential to enhance researchers’ effectiveness and the accessibility of scientific documents.