MRGSEM-Sum: An Unsupervised Multi-document Summarization Framework based on Multi-Relational Graphs and Structural Entropy Minimization

TL;DR

MRGSEM-Sum is an unsupervised multi-document summarization framework that models complex relationships with multi-relational graphs and uses structural entropy minimization for adaptive sentence clustering, producing high-quality summaries.

Contribution

It introduces a novel multi-relational graph construction and an entropy-based clustering method that automatically determines the optimal number of sentence groups for summarization.

Findings

Outperforms previous unsupervised methods on benchmark datasets.

Achieves comparable performance to supervised models and large language models.

Human evaluation shows high summary consistency and coverage.

Abstract

The core challenge faced by multi-document summarization is the complexity of relationships among documents and the presence of information redundancy. Graph clustering is an effective paradigm for addressing this issue, as it models the complex relationships among documents using graph structures and reduces information redundancy through clustering, achieving significant research progress. However, existing methods often only consider single-relational graphs and require a predefined number of clusters, which hinders their ability to fully represent rich relational information and adaptively partition sentence groups to reduce redundancy. To overcome these limitations, we propose MRGSEM-Sum, an unsupervised multi-document summarization framework based on multi-relational graphs and structural entropy minimization. Specifically, we construct a multi-relational graph that integrates semantic and discourse relations between sentences, comprehensively modeling the intricate and dynamic connections among sentences across documents. We then apply a two-dimensional structural entropy minimization algorithm for clustering, automatically determining the optimal number of clusters and effectively organizing sentences into coherent groups. Finally, we introduce a position-aware compression mechanism to distill each cluster, generating concise and informative summaries. Extensive experiments on four benchmark datasets (Multi-News, DUC-2004, PubMed, and WikiSum) demonstrate that our approach consistently outperforms previous unsupervised methods and, in several cases, achieves performance comparable to supervised models and large language models. Human evaluation demonstrates that the summaries generated by MRGSEM-Sum exhibit high consistency and coverage, approaching human-level quality.