Which Information Matters? Dissecting Human-written Multi-document Summaries with Partial Information Decomposition

Understanding the nature of high-quality sum maries is crucial to further improve the perfor mance of multi-document summarization. We propose an approach to characterize human written summaries using partial information decomposition, which decomposes the mutual information provided by all source documents into union, redundancy, synergy, and unique in formation. Our empirical analysis on different MDS datasets shows that there is a direct de pendency between the number of sources and their contribution to the summary.