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

TL;DR

This paper introduces a method to analyze human-written multi-document summaries by decomposing the information they contain into components like redundancy and synergy, revealing how source documents contribute to the summary.

Contribution

It applies partial information decomposition to characterize the informational structure of summaries, offering new insights into the role of multiple sources in summarization.

Findings

Number of sources correlates with their contribution to summaries.

Redundancy and synergy play significant roles in information composition.

Empirical analysis across datasets validates the approach.

Abstract

Understanding the nature of high-quality summaries is crucial to further improve the performance 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 information. Our empirical analysis on different MDS datasets shows that there is a direct dependency between the number of sources and their contribution to the summary.