Introduction
Automatic text summarization hinges on the ability to identify significant content within a document. Traditional methods often rely on frequency-based metrics like TF-IDF. However, these metrics fail to account for the structural and relational importance of concepts within a text. This study explores the application of social network analysis—specifically centrality measures—to identify key noun phrases (NPs) that are likely to appear in human-written abstracts. By modeling a document as a network of interacting NPs, we can identify "influential" concepts that define the document's essence.
Methodology
Network Construction
The proposed pipeline begins with the extraction of noun phrases using a POS tagger and a set of 35 parsing rules. These NPs serve as nodes in a graph. Edges are established based on the sequential appearance of NPs in the text. Two scales of network construction were evaluated:
- Sentence Level: Links are restricted to NPs appearing within the same sentence.
- Document Level: Links are created between all sequential NPs throughout the entire document.
Centrality Measures
Four centrality measures were employed to rank node importance:
- Degree Centrality (DC): Measures local connectivity (number of direct neighbors).
- Closeness Centrality (CC): Measures the average distance to all other nodes.
- Betweenness Centrality (BC): Measures how often a node lies on the shortest path between others.
- Information Centrality (IC): Accounts for information flow across all possible paths.
Prediction Models
Two approaches were tested for predicting abstract-worthy NPs:
- Heuristic Approach: Selecting the top $N$ NPs based solely on centrality rankings.
- Machine Learning Approach: Utilizing the YaDT (Yet another Decision Tree) algorithm. Centrality measures were integrated as features alongside frequency, document position, and linguistic markers.
Experimental Results and Discussion
The study utilized the CMP-LG Corpus, consisting of 178 scientific documents with human-written abstracts.
Heuristic Performance and Network Topology
At the sentence level, the network often fragments into disconnected subgraphs. As shown in Figure 4, this fragmentation significantly impacts global measures like Closeness and Betweenness centrality. In these sparse networks, Degree Centrality—a local measure—emerges as the most effective heuristic.
When the network is expanded to the document level, it typically forms a single connected component, causing the performance of the four centrality measures to converge.
Machine Learning Integration
The most significant performance gains were observed when centrality measures were used as features in a supervised learning model. While standalone centrality heuristics yielded F-measures around 0.25, the Decision Tree model achieved a substantial improvement.
As illustrated in Figure 5, the ensemble of features (including centrality, frequency, and position) allows the model to achieve a Precision of ~0.81 and a Recall of ~0.97, resulting in an F-measure of 0.88. Statistical analysis confirmed that the inclusion of centrality measures provides a significant improvement over models relying solely on frequency or position.
Conclusion
This research demonstrates that centrality measures provide a valuable signal for identifying significant concepts in scientific text. While network topology influences the effectiveness of specific measures, their integration into machine learning models offers a robust framework for predicting the content of human-written abstracts. Future work will explore weighted edges and coreference resolution to further refine the network representation.
Summary updated