Abstract
Vahed Qazvinian, Dragomir R. Radev. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics.
2010.
Introduction
Automated citation-based summarization aims to generate literature surveys by aggregating how a target paper is cited across the scientific community. However, traditional methods often rely solely on "explicit" citing sentences—those containing formal citation markers (e.g., "Eisner (1996)"). This approach frequently misses the "meat" of the citation: the descriptive context, algorithm details, or specific findings that typically reside in the surrounding sentences. These "non-explicit" citing sentences are crucial for creating rich, informative summaries. The challenge lies in accurately identifying which neighboring sentences continue the discussion of the cited work and which transition to new topics.
Methodology: Markov Random Fields for Sequence Labeling
The authors frame the identification of citation context as a sequence labeling task. Rather than using a simple independent classifier, they employ a Markov Random Field (MRF) to capture the dependencies between adjacent sentences.
The Graph Model and Potential Functions
In this model, each sentence in a document is represented as a node. The state of a node (whether it is a context sentence or not) is influenced by its neighbors through potential functions. The model incorporates three primary signals: explicit citations, lexical similarity to the cited work, and discourse cues (e.g., "This approach").
Belief Propagation (BP)
To infer the most likely labels for all sentences, the authors use Belief Propagation. This involves passing "messages" between nodes until the system reaches a stable state of "beliefs" about each sentence's label.
Experimental Setup and Neighborhood Structures
The researchers evaluated different neighborhood configurations to determine the optimal scope of citation context:
- BP1: A linear chain where each sentence only interacts with its immediate predecessor and successor.
- BP4: A local neighborhood where each sentence is influenced by four sentences on either side.
- BPn: A fully connected graph where every sentence in the paper is connected to every other sentence.
The study utilized the ACL Anthology Network (AAN), with 10 source papers and 203 distinct paper-reference pairs. Expert annotations served as the gold standard, showing high inter-judge agreement (Kappa > 0.85).
Key Results and Insights
The experiments revealed several critical findings regarding how authors cite work:
- The Proximity Rule: Most context sentences are clustered tightly around the explicit citation. While "gaps" (sentences digressing from the topic) occur, they are relatively rare.
- Optimal Context Window: The BP4 model consistently outperformed other configurations. This suggests that citation context is a local phenomenon, but one that typically spans more than just the immediate neighboring sentence.
- Local vs. Global Noise: The fully connected BPn model performed poorly, confirming that connecting distant sentences introduces significant noise rather than useful global context.
Performance Metrics
The MRF-based BP4 approach achieved an $F_{\beta=3}$ score (prioritizing recall) of up to 0.889, significantly exceeding similarity-based baselines. Furthermore, when these extracted sentences were used to generate summaries via LexRank, the "Pyramid scores" (a measure of summary quality) improved from 0.416 (explicit only) to 0.634 (with context).
Conclusion
By automatically identifying non-explicit citing sentences using an MRF framework, this research significantly enhances the depth and quality of citation-based summarization. While the current system excels at extraction, future work is needed to improve the narrative fluency of the generated surveys and to potentially incorporate global bibliometric data to further refine sentence selection.
Summary updated