DOI: 10.35377/saucis...1853948 ISSN: 2636-8129

Energy-Based Interaction-Aware Graph Modeling for Unsupervised Long-Document Extractive Summarization

Abubakar Bashir, Abdulkadir Abubakar Bichi
Show PDF Cite
Long-document extractive summarisation presents persistent challenges, including high redundancy, diffuse topical structure, and limited availability of high-quality supervision. Existing unsupervised graph-based methods improve upon classical centrality algorithms but rely on iterative ranking procedures that are sensitive to graph density and lack formal optimality guarantees. This paper introduces a fully unsupervised extractive summarisation framework in which sentence salience is formulated as a convex energy minimization problem over a semantic-similarity graph. The objective combines a centrality-driven prior with graph Laplacian regularisation, enabling joint inference of sentence importance while preserving convexity and admitting a closed-form solution via a single linear system solve. Discrete sentence selection is performed using a salience threshold, followed by a cosine-similarity-based redundancy filter in a separate greedy stage, preserving the tractability and global optimality of the inference step. Experiments on GovReport, BillSum, and PubMed demonstrate consistent improvements in ROUGE-1 over classical graph-based baselines, indicating improved unigram content coverage. Performance on ROUGE-2 and ROUGE-L is competitive with recent unsupervised approaches on some datasets but shows gaps on others, most notably in ROUGE-L on BillSum, which we attribute to the redundancy filter's effect on local sentence-level coherence. Sensitivity analyses confirm stability across a broad range of graph regularisation strengths and sparsity thresholds. These results support convex energy-based modeling as a principled, reproducible, and domain-independent alternative to heuristic iterative ranking for unsupervised long-document summarisation, while also identifying local coherence preservation as a direction for future improvement.

More from our Archive