Abstract
The ZJU-EDL system addresses the complexities of Entity Discovery and Linking (EDL) through a cascaded neural architecture and a novel recursive denoising strategy. By utilizing BERT-based models and a context-focused cross-validation technique to clean distantly supervised data, the system achieved competitive performance in the TAC KBP 2019 evaluation.
Introduction
Entity Discovery and Linking (EDL) involves identifying entity mentions in text and linking them to a knowledge base or assigning fine-grained types. The ZJU-EDL system specifically targets two major hurdles: significant noise in distantly supervised training data and the challenge of mapping entities to the complex AIDA taxonomy required by the TAC KBP competition.
System Architecture
The ZJU-EDL system employs a two-stage cascaded pipeline consisting of a Location Model and a Typing Model, followed by a document-level rectification step.
Location Model (Mention Detection)
This component identifies entity boundaries using a BERT backbone for feature extraction, followed by Bi-LSTM and CRF (Conditional Random Field) layers for I-O-B tagging. To improve robustness, the system integrates the spaCy NLP library, leveraging its high precision to refine the boundaries detected by the neural model.
Typing Model
Once mentions are located, the Typing Model assigns specific categories. It uses BERT to represent both the mention and its context. To manage fine-grained types, the model incorporates attention mechanisms and hierarchy-aware loss functions. Furthermore, a Knowledge Base (KB) acts as a filter; if a mention links to a known entity, the system restricts possible types to those permitted by the KB, significantly reducing the search space.
Recursive Denoising Procedure
A core innovation of the ZJU-EDL system is its recursive denoising strategy, designed to clean noisy training data generated via distant supervision.
The dataset is split into $k$-folds. During training, the model is trained on $k-1$ folds while the mention string itself is masked in the validation fold. This forces the model to predict the entity type based solely on context, preventing it from simply memorizing noisy labels associated with specific words. If the model predicts a different label with high confidence, the original label is updated. This iterative process "polishes" the dataset, significantly reducing the impact of distant supervision errors.
Datasets and Evaluation
The training corpus consists of 8 million instances merged from four major sources: Figer, Ultra-fine Entity Typing, Wikipedia, and the RPI silver-standard corpus.
In the TAC KBP 2019 EDL track, the system was evaluated across two windows. Window 2, which included document-level rectification (selecting the highest-scoring type for a mention across an entire document), showed a notable performance increase:
- Window 1: Score: 0.395 (Precision: 0.375, Recall: 0.418)
- Window 2: Score: 0.446 (Precision: 0.423, Recall: 0.472)
Conclusion
The ZJU-EDL system demonstrates that context-driven denoising and hybrid mention detection are effective strategies for large-scale EDL tasks. While the system achieved competitive results, future work will focus on improving the handling of overlapping types and further refining the denoising of massive, high-noise datasets like RPI.
Summary updated