A Knowledge Graph‐Guided Hybrid Extractive–Abstractive Framework for Multi‐Document Summarization Using Encoder–Decoder Language Models

ABSTRACT Multi‐document summarization attempts to compress information from multiple related documents, maintaining fact and language fluency. Existing extractive summarization methods effectively maintain factual correctness but often lack originality and coherence, whereas abstractive approaches generate fluent summaries but struggle with redundancy and factual inconsistency, particularly when multiple documents contain overlapping or related information. In this case, the gaps in previous work are overcome by using a knowledge graph hybrid extractive and abstractive summarization model. The model performs text preprocessing techniques, including text cleaning, sentence boundary detection, tokenization, lemmatization, and stop word removal. Entities and semantic relations are subsequently extracted using Named Entity Recognition, the ReBEL model, and a fallback Subject–Verb–Object parsing strategy. The extracted relations are fed into a graph creation algorithm, and the graph is weighted. This graph is filtered to retain salient edges, which are then used to create batches of factual information. These fact‐batches serve as structured prompts to local encoder–decoder language models, such as FLAN‐T5, fine‐tuned FLAN‐T5 with LoRA, and BART. Among the three models, the proposed approach achieves the best performance, obtaining ROUGE‐1, ROUGE‐2, ROUGE‐L and BERTS score, KG‐based factual consistency of 0.446, 0.153, and 0.427, 0.88, 0.84, respectively. Therefore, the proposed model outperforms the base FLAN‐T5 and BART models in terms of summary quality. The results confirm that integrating KG‐based factual representations with encoder–decoder language models significantly enhances factual accuracy and overall summary quality.