From Extraction to Reasoning: A Systematic Review of Algorithms in Multi-Document Summarization and QA

Abstract

Multi-document summarization and question-answering (QA) have become pivotal tasks in Natural Language Processing (NLP), facilitating information extraction and decision-making across various domains. This systematic review explores the evolution of algorithms used in these tasks, providing a comprehensive taxonomy of traditional, modern, and emerging approaches. We examine the progression from early extractive methods such as TF-IDF and TextRank, to the advent of neural models like BERT, GPT, and T5, and the integration of retrieval-augmented generation (RAG) for QA. Hybrid models combining traditional techniques with neural approaches and graph-based methods are also discussed. Through a detailed analysis of algorithmic frameworks, we identify key strengths, weaknesses, and challenges in current methodologies. Additionally, the review highlights recent trends such as unified models, multimodal algorithms, and the application of reinforcement learning in summarization and QA tasks. We also explore the real-world relevance of these algorithms in sectors such as news, legal, medical, and education. The paper concludes by outlining open research directions, proposing new evaluation frameworks, and emphasizing the need for cross-task annotations and ethical considerations in future algorithmic development.