UM-MDAS: A Unified Model Framework for Multi-Document Abstractive Summarization

The volume of information across diverse sources has been increasing rapidly, and thus, there has been a need for effective multi-document summarization systems. Multi-Document Abstractive Summarization (MDAS) aims to generate concise and coherent summaries that capture common core ideas between related documents. Despite recent improvements in transformer-based models in the transformer-based models, the current methodologies of MDAS still struggle with cross-document redundancy, inconsistent content selection, and the inability to control redundancy in the salience of a summary. The article introduces a framework, UM-MDAS, which is a unified and modular architecture of multi-document abstractive summarization to overcome these challenges with a structured pipeline architecture. The framework builds sentence-level semantic representations based on SBERT and combines the representations into document-level representations through min-max pooling to identify salient and extreme semantic features. The cosine similarity is used to measure semantic relevance and eliminate redundancy across document clusters, whereas the abstractive summary generation is done through the PEGASUS sequence-to-sequence model. The experiments carried out with MultiNews data indicate that UM-MDAS enhances semantic coverage, fluency, and coherence. Evaluation results show ROUGE-1 (53.83), ROUGE-2 (21.00), ROUGE-L (23.54), and BERTScore (86.41), indicating that the proposed framework performs well in capturing both surface-level and contextual similarities. The article introduces a framework, UM-MDAS, which is a unified and modular architecture of multi-document abstractive summarization to overcome these challenges with a structured pipeline architecture.