Multi-Document Summarization

Extractive methods select and concatenate the most important sentences directly from the source documents, preserving the original wording. LexRank is a prominent example. Abstractive methods generate new text that paraphrases and fuses content from multiple sources, producing more fluent summaries but requiring pretrained language models and greater computational resources.

The method is designed for clusters of related documents — at least five is a practical minimum. Below that threshold, a standard single-document summariser is more appropriate. There is no strict upper limit, but very large clusters may require hierarchical or incremental strategies to manage computational cost.

The standard automatic metric is ROUGE, which counts n-gram overlaps between the generated summary and one or more human reference summaries. However, ROUGE correlates imperfectly with human judgements of coherence and informativeness, so human evaluation remains the gold standard in research and production contexts.

Contradiction handling is one of the hardest open problems in MDS. Most off-the-shelf methods do not explicitly detect conflicts — they may silently favour one perspective or produce an internally inconsistent summary. For domains where contradictions matter (e.g., scientific or legal text), explicit contradiction detection should be added as a preprocessing step before summarisation.