Hierarchical Graph Convolutional Networks for Jointly Resolving Cross-document Coreference of Entity and Event Mentions

Duy Phung, Tuan Ngo Nguyen, Thien Huu NguyenNAACL 2021

TL;DR

This paper introduces a novel hierarchical Graph Convolutional Network (GCN) model for jointly resolving cross-document event and entity coreference. By leveraging pruned dependency trees for sentence-level context and interaction graphs for document-level context, the model achieves state-of-the-art performance on the ECB+ dataset.

What Problem Does This Paper Address?

The paper tackles the challenging problem of Cross-Document Event Coreference Resolution (CDECR). CDECR aims to identify mentions of events across different documents that refer to the same real-world event. A key difficulty lies in effectively capturing the relationships between event mentions and their arguments (participants, locations, times), as these often provide crucial clues for coreference. Prior work has struggled to fully leverage this argument-related information and has often relied on simpler lexical features for document representation, limiting their ability to generalize and capture deeper semantic meaning.

Methods

The authors propose a novel hierarchical GCN model that jointly resolves entity and event coreference. The core of their approach involves two levels of GCNs:

  • Sentence-Level GCNs: These GCNs operate on pruned dependency trees. The pruning process focuses on the shortest dependency paths between event triggers and their arguments, along with important context words. This allows the model to build richer representations for individual event and entity mentions by explicitly encoding relevant contextual information. They use separate GCNs for entities and events, with a regularization loss to encourage similar representations, promoting knowledge transfer.
  • Document-Level GCN: This GCN operates on interaction graphs constructed from entity and event mentions within a document. These graphs emphasize predicate-argument relationships and entity coreference links. The sentence-level GCN representations serve as input to this document-level GCN, enabling the model to capture document-wide context and generate document representations.

The model is trained iteratively, alternating between clustering entity mentions and event mentions. This joint approach allows the two tasks to mutually benefit from each other.

Datasets

The primary dataset used for evaluation is the ECB+ dataset. This dataset is noted as the largest available with both within- and cross-document annotations for entity and event coreference. The authors follow a specific setup and split for this dataset, consistent with prior work, which utilizes a validated subset and focuses on gold-standard mentions for evaluation.

Key Insights and Ideas

  • Pruned Dependency Trees for Context: The idea of using pruned dependency trees, focusing on shortest paths between triggers and arguments, is crucial. This allows the model to selectively capture the most relevant contextual words, rather than being overwhelmed by the entire sentence structure.
  • Hierarchical GCNs for Multi-Level Context: The combination of sentence-level GCNs (capturing local context) and document-level GCNs (capturing global context) provides a powerful way to represent mentions at different granularities.
  • Joint Entity and Event Resolution: The paper reinforces the benefit of jointly modeling entity and event coreference, showing how this interdependence can lead to improved performance for both tasks.
  • Interaction Graphs for Document Representation: Using interaction graphs that highlight entity-event relationships and entity coreference within a document is a novel way to generate document representations, moving beyond simple lexical features.

Key Results

The proposed hierarchical GCN model (HGCN) achieves state-of-the-art performance on the ECB+ dataset for both cross-document event and entity coreference resolution. Specifically, HGCN achieved a CoNLL F1 score of 81.3% for event coreference, outperforming the prior SOTA model MCS by 1.3%, and a CoNLL F1 score of 72.4% for entity coreference, surpassing BSE by 1.2%. These improvements are statistically significant (p < 0.001). Ablation studies demonstrate that each component of the proposed model (pruned dependency trees, separate sentence-level GCNs, GCN interaction, and document-level GCN) contributes significantly to the overall performance.

Limitations and Future Work

The authors acknowledge that their model's performance is sensitive to the quality of the underlying ELMo embeddings. They also suggest that future work could explore better mechanisms for identifying important context words for CDECR. Additionally, they mention investigating the distribution of error types to guide future improvements.

Summary updated

BibTeX
@inproceedings{phung2021hierarchical,
  author = {Duy Phung and Tuan Ngo Nguyen and Thien Huu Nguyen},
  title = {{Hierarchical Graph Convolutional Networks for Jointly Resolving Cross-document Coreference of Entity and Event Mentions}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure