BEL: Bagging for Entity Linking

Zhe Zuo, Gjergji Kasneci, Toni Grütze, Felix NaumannCOLING 2014 · OpenReview

Abstract: Zhe Zuo, Gjergji Kasneci, Toni Gruetze, Felix Naumann. Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers.

2014.

Introduction

Named-Entity Linking (NEL) is the critical task of mapping text mentions to unique entities in a knowledge base (KB). While essential for information extraction, existing NEL systems often struggle with context sensitivity, explicit handling of uncertainty, and computational efficiency. High-accuracy "joint reasoning" models, which ensure semantic coherence across all mentions in a document, are frequently too slow for web-scale applications. The BEL (Bagging for Entity Linking) framework addresses these challenges by employing an ensemble of simple classifiers, achieving state-of-the-art accuracy with significantly reduced latency.

The BEL Algorithm

The core of the BEL framework is an ensemble approach that treats entity disambiguation as a majority-voting problem. The process consists of four main stages:

  1. Candidate Pruning: To maintain efficiency, BEL uses a "prominence score" derived from Wikipedia link frequencies to narrow the potential entities for a mention to the top 40 candidates.
  2. Bagging (Bootstrap Aggregating): Instead of a single pass over the context, BEL identifies a "range of relevant terms" (optimized to 55 terms) and creates multiple random sub-samples from this range with replacement.
  3. Ensemble Classification: For each sub-sample, a simple language-model-based classifier ranks the candidate entities. The final decision is determined by a majority vote across the ensemble (typically 199 subsets).
  4. Local Reasoning: BEL treats mentions independently but exploits document-level repetition. If a mention appears multiple times, the system reuses previous reliable links to optimize performance.
Overview of the BEL algorithm showing the bagging process and ensemble classification.

Experimental Setup

BEL was evaluated against three diverse benchmarks using YAGO2 as the knowledge base:

  • CoNLL-YAGO: 76 Reuters news articles (1,431 mentions).
  • CUCERZAN: 336 Wikipedia articles (5,343 mentions).
  • KORE: 50 short, highly ambiguous sentences (148 mentions) designed to test sparse context performance.

Analysis and Insights

The Impact of Bagging and Pruning

A key finding is that textual context is often noisy. By randomly sampling subsets of terms, the ensemble becomes robust to irrelevant words. Bagging improved precision by approximately 2.5% compared to a single classifier. Furthermore, the "prominence" prior proved highly effective; the correct entity resides within the top 40 candidates 92.4% of the time, justifying aggressive pruning.

Impact of candidate pruning on coverage (a) and number of sub-samples on precision (b).

Performance and Efficiency

BEL achieved an F1-score of 87.98% on the CoNLL-YAGO dataset, outperforming the complex AIDA-GRAPH model (86.59%). More importantly, BEL is orders of magnitude faster. On the CUCERZAN dataset, BEL completed processing in approximately 11 minutes, whereas the AIDA-KORE model required over 25 hours. This efficiency stems from the independent nature of the ensemble classifiers, which makes the algorithm inherently parallelizable.

Conclusion

BEL demonstrates that complex joint reasoning is not always necessary for high-quality entity linking. By leveraging local context through bagging, the system achieves superior accuracy and scalability. While it faces challenges in extremely sparse contexts (as seen in the KORE dataset) and requires more robust NIL-detection for non-KB entities, BEL represents a significant step toward efficient, web-scale entity linking.

Summary updated

BibTeX
@inproceedings{zuo2014bel,
  author = {Zhe Zuo and Gjergji Kasneci and Toni Grütze and Felix Naumann},
  title = {{BEL: Bagging for Entity Linking}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2014},
  url = {https://openreview.net/forum?id=xnzYkpd057},
}
Figure