Abstract
This paper presents a novel approach to single-document extractive summarization using dense semantic graphs. By leveraging dependency tree proximity between nouns and adjectives rather than traditional Subject-Verb-Object (SVO) triplets, the method captures a richer topical structure. Experimental results on DUC-01 and DUC-02 datasets demonstrate that increasing graph density through grammatical distance thresholds significantly improves sentence ranking and summarization performance compared to sparse triplet-based models.
Introduction
In extractive summarization, the primary challenge is identifying the most salient sentences within a document to form a concise summary. Graph-based representations are frequently employed for this task, where nodes represent linguistic units and edges represent their relationships. However, traditional "semantic graphs" often rely on logical triples (Subject-Verb-Object). While SVO triples capture core actions, they frequently discard crucial context. For instance, in the sentence "President Obama’s arrival in London created a joyful atmosphere," a triple-based system might only extract Arrival -> create -> atmosphere, losing the vital connections to "London" and "Obama."
Furthermore, if an important entity is not a subject or object in a specific sentence, it may remain unlinked, resulting in a sparse graph that fails to reflect the document's true topical density. This paper addresses these gaps by proposing a method to construct a dense semantic graph that preserves thematic weight.
Methodology
The proposed method constructs a Dense Semantic Graph through a multi-step pipeline:
- Preprocessing: Stanford CoreNLP is utilized for co-reference and pronominal resolution, ensuring that different mentions of the same entity (e.g., "he" or "she" vs. a proper name) are linked to a single node.
- Node Selection: The graph focuses on Nouns and Adjectives, as these parts of speech typically carry the most thematic and descriptive weight in a document.
- Edge Creation: Instead of limiting connections to SVO components, edges are created based on the dependency tree of each sentence. An edge is formed between two nodes if the shortest path between them in the tree is below a specific threshold (varied from 2 to 5 in testing).
- Ranking and Scoring: The PageRank algorithm is applied to the resulting dense graph to determine word importance based on connectivity. Sentence scores are then calculated as the sum of the PageRank scores of their constituent nodes.
Results and Discussion
The researchers evaluated the approach using the DUC-01 and DUC-02 benchmarks, comparing system-generated summaries against human "gold standards" using ROUGE-1 scores.
Key findings include:
- Density vs. Sparsity: On the DUC-01 dataset, the dense semantic graph (SD-5) achieved a ROUGE-1 score of 0.32974, a significant improvement over the 0.31793 achieved by the basic triplet-based system.
- The "Distance 5" Threshold: Performance improved as the allowed distance between words in the dependency tree increased, peaking at a distance of 5. Beyond this, scores plateaued, suggesting that more distant grammatical relations introduce noise rather than useful context.
- Positional Heuristics: Combining the dense graph with a sentence position boost (favoring earlier sentences) further increased the ROUGE-1 score to 0.33753.
While the system was slightly outperformed by the Open Text Summarizer (OTS) (0.35134), the authors note that OTS relies on extensive language-specific lexicons, whereas the dense semantic graph approach is purely structural and algorithmic.
Conclusion
The dense semantic graph approach demonstrates that grammatical structure is a powerful proxy for thematic importance. By moving beyond simple triples to a denser representation of nouns and adjectives, the model more accurately identifies central themes. Future work will focus on incorporating lexical knowledge and word sense disambiguation to connect conceptually related but linguistically distinct nodes.
Summary updated