Transforming Multi-Document Abstractive Summarization: A Comparative Study of State-of-the-Art Models and Benchmarks

This study presents a comprehensive comparative analysis of state-of-the-art models for multi-document abstractive summarization, focusing on their performance across key datasets, including MultiNews, WikiSum, DUC, and CNN/Daily Mail. Given the rapid proliferation of text data, the ability to generate coherent and succinct summaries from multiple documents holds immense potential for knowledge dissemination, research acceleration, and efficient information retrieval. To achieve this goal, we meticulously evaluated prominent models such as BART, Pegasus, T5, Longformer, Pointer-Generator Networks, and more. Extensive experimentation involved fine-tuning each model for optimal performance using benchmark datasets, measuring their effectiveness through ROUGE-1, ROUGE-2, ROUGE-L scores, and additional precision metrics. Results indicate that PRIMERA outperformed its counterparts, achieving the highest ROUGE scores across the datasets, demonstrating superior summarization quality and coherence. The findings underscore the nuanced capabilities of transformer-based architectures compared to hierarchical and graph-based approaches, offering insights into scalability and generalization for multi-document tasks. This research reflects rigorous data processing and performance tuning, leveraging a custom dataset creation pipeline and deep analysis of generated summaries. Our work advances the field by identifying optimal strategies and architectures for multi-document summarization, paving the way for future innovations and practical applications in natural language processing.