Intra-document and Inter-document Redundancy in Multi-document Summarization

Multi-document summarization differs from single-document summarization in excessive redundancy of mentions of some events or ideas. We show how the amount of redundancy in a document collection can be used for assigning importance to sentences in multi-document extractive summarization: for instance, an idea could be important if it is redundant across documents because of its popularity; on the other hand, an idea could be important if it is not redundant across documents because of its novelty. We propose an unsupervised graph-based technique that, based on proper similarity measures, allows us to experiment with intra-document and inter-document redundancy. Our experiments on DUC corpora show promising results.