Hybrid extractive–abstractive summarization of scientific texts via deep clustering and attention mechanism

Scientific document summarization presents unique challenges due to domain-specific terminology, long-form discourse, and high compression demands. We propose a novel hybrid summarization model that combines deep clustering-based extractive scoring with an attention-guided abstractive generator. Sentence embeddings are clustered to capture semantic structure, and their importance is estimated via an MLP-based scoring system augmented with cluster-level weighting. These representations are then passed to a cross-attention-driven transformer-based decoder, enabling contextualized summary generation. Evaluated on ArXiv and PubMed datasets, our model surpasses BART, PEGASUS, and T5 in both ROUGE and BLEU metrics. Ablation studies confirm the critical role of clustering, scoring, and fusion components. Our approach bridges extractive precision and abstractive richness, demonstrating promising applicability in assisting researchers with scientific information overload.