<![CDATA[Decision-making processes often need summaries that are concise, accurate, and easy to verify. However, neural abstractive summarization models may produce unsupported information or hallucinations. This study explores citation-constrained multi-document summarization using the Multi-News dataset, where each important statement in the summary is connected to a supporting sentence through citation markers such as [S1] and [S2]. The proposed method applies an evidence-grounded pipeline consisting of sentence segmentation, Top-K evidence retrieval using BM25 or dense embedding methods, and a BART-based model for generating citation-aware summaries. Since the Multi-News dataset does not provide citation annotations, silver-standard citation labels are created using TF-IDF cosine similarity. An optional alignment filtering process is also applied to reduce weakly supported content. The evaluation process includes ROUGE scores, citation precision and coverage, as well as consistency checks for entities and numerical information. Experimental results on a small reproducible dataset show that the proposed approach can generate summaries with measurable citation quality while maintaining ROUGE performance comparable to non-citation models. Overall, this framework improves transparency and supports more reliable and auditable summarization under limited computational resources.]]>
Copyrights © 2025
