Abstract: Xiaojun Wan. Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing.
2008.
Introduction
Multi-document summarization (MDS) is the process of generating a concise, representative summary from a collection of documents focused on the same topic. A dominant paradigm in this field has been the graph-based ranking algorithm, exemplified by methods like LexPageRank. These methods typically represent sentences as nodes in a graph, with edges representing semantic similarity. Sentences that receive the most "votes" from other similar sentences are deemed highly salient and selected for the summary.
However, traditional graph-based models often operate under a "flat" assumption, treating all sentences as independent entities regardless of their source document. This approach ignores critical document-level context. In a typical document cluster, some articles are central to the topic while others are peripheral. Furthermore, the internal structure of a document—such as the tendency for lead sentences to carry more weight—provides a significant signal for importance that flat models fail to capture.
The Document-Based Graph Model (DGM)
To address these limitations, this paper introduces the Document-Based Graph Model (DGM). Unlike the standard flat graph, DGM utilizes a two-layer hierarchical structure consisting of a sentence layer and a document layer. This architecture allows the model to incorporate document-level metadata into the sentence ranking process.
The core innovation of DGM is the conditional affinity weight. When determining the strength of a relationship between two sentences ($v_i$ and $v_j$), the model weights their cosine similarity based on two primary factors:
- Document Importance ($\pi$): This metric reflects how central a document is to the entire cluster. The author evaluated three methods for calculation: simple cosine similarity to the cluster centroid, PageRank on a document-level graph, and a uniform distribution baseline.
- Sentence-Document Correlation ($\omega$): This measures how well a sentence represents its parent document. The study tested both content-based similarity (cosine similarity to the document) and position-based heuristics (assigning higher weights to lead sentences).
By integrating these factors, the influence or "vote" one sentence exerts on another is amplified if the sentences originate from prestigious documents or occupy highly representative positions within those documents.
Experimental Setup
The DGM was evaluated using two benchmark datasets from the Document Understanding Conference:
- DUC 2001: 30 document sets containing 309 documents.
- DUC 2002: 59 document sets containing 567 documents.
Both datasets consist of news articles from TREC-9. The performance was measured using ROUGE metrics (ROUGE-1, ROUGE-2, and ROUGE-W) against human-written 100-word summaries.
Key Results and Discussion
The experimental results demonstrated that DGM variants consistently outperformed the standard Graph Model (GM) baseline across all metrics. A key finding was the robustness of the hierarchical approach; even simple methods for calculating document importance yielded improvements over the flat model.
As shown in the performance analysis, the parameter $\lambda$—which balances the influence of the two documents involved in a sentence-to-sentence link—plays a role in optimization, though most DGM configurations remained superior to the GM baseline (represented by the solid black line) across a wide range of $\lambda$ values. Notably, the position-based correlation factor ($\omega$) performed exceptionally well, confirming the strong "lead bias" inherent in news journalism. On the DUC 2001 dataset, several DGM configurations even surpassed the top-performing systems from the original competition.
Conclusion and Future Work
This research highlights that a sentence's "pedigree"—its document-level context—is a powerful indicator of its saliency. By moving from a flat graph to a hierarchical model, DGM captures the inherent structure of document clusters more effectively than traditional methods.
Future research directions include increasing the granularity of the model by adding a "sub-topic" layer using algorithms like TextTiling. Additionally, exploring how these document-level signals behave in non-news domains, such as legal or scientific literature, remains an open and promising area of study.
Summary updated