Abstract
Heng Ji, Zheng Chen. Proceedings of Human Language Technologies: The 2009 Annual Conference of the North American Chapter of the Association for Computational Linguistics, Companion Volume: Demonstration Session. 2009.
This paper presents a system that moves beyond extracting isolated facts from single documents to tracking key individuals across entire news corpora. By identifying "centroid arguments"—the most salient people in a collection—the system automatically constructs coherent event chains, mapping a person's activities onto both a temporal timeline and a geographic map.
Introduction
Traditional Information Extraction (IE) systems are often "short-sighted," typically processing text one sentence or one document at a time. This results in massive databases filled with unconnected, redundant, and sometimes contradictory facts. For a human analyst, trying to piece together a person's movements or an organization's history from these fragmented outputs is a manual, time-consuming nightmare.
The authors address the lack of discourse coherence in IE. They argue that by focusing on the most central entities (the "protagonists" of the news), we can bridge conceptual gaps and transform a pile of isolated facts into a structured, navigable narrative.
System Architecture and Methods
The system follows a multi-stage pipeline to turn raw text into a visual "story":
- Within-document IE: Using a state-of-the-art baseline, the system extracts entities (people, locations, etc.), time expressions, and events based on the 33 types defined in the ACE05 standard. It uses Maximum Entropy classifiers to determine event types and participant roles.
- Centroid Argument Detection: Instead of treating all names equally, the system calculates a Global Confidence score. It retrieves topically related documents using the INDRI search engine and looks for arguments that appear frequently and with high local extraction confidence. These high-scoring individuals are labeled "centroid arguments."
- Cross-document Event Aggregation: To handle redundancy, the system merges events involving the same centroid argument using four specific operations: Coreference, Subset, Subsumption, and Complement.
Temporal and Spatial Tracking
Finally, the system normalizes dates and locations to plot the event chain on a synchronized timeline and map interface. This allows for a synthesized view of an individual's history across thousands of documents.
Key Insights and Results
- Salience as a Filter for Accuracy: One of the most important findings is that "globally salient" arguments are much more likely to be extracted correctly. By ranking results by global confidence, the system effectively filters out extraction errors. The raw IE system had an accuracy of about 53%, but the top 85 ranked arguments achieved over 70% accuracy.
- Coherence through Centering: Borrowing from linguistic "Centering Theory," the authors show that tracking a single central entity provides a natural "thread" that makes a complex web of events easier for humans to understand.
- Scalability: The system identified 7,962 centroid arguments across the English TDT5 corpus. While the offline processing took about three hours, the front-end browsing experience is near-instant (one second).
Conclusion and Future Work
The current demonstration focuses primarily on person entities. The authors acknowledge that extending this to organizations (e.g., tracking company mergers and acquisitions) is a logical next step. Additionally, they plan to allow users to adjust "compression ratios" to see more or less detail depending on their analytical needs. This system moves the needle from simple Information Retrieval toward true Knowledge Synthesis.
Summary updated