Cross-Document Event Coreference Resolution on Discourse Structure

Xinyu Chen, Sheng Xu, Peifeng Li, Qiaoming ZhuEMNLP 2023 · OpenReview

Abstract

Cross-document event coreference resolution (CD-ECR) is the task of clustering event mentions across multiple documents that refer to the same real-world events. Previous studies usually model the CD-ECR task as a pairwise similarity comparison problem using local event mention features, often ignoring implicit global information and failing to capture interactions between long-distance mentions. To address this, we incorporate discourse structure as global information. We utilize a discourse rhetorical structure constructor to build tree structures representing documents, then extract shortest dependency paths from these trees to represent interactions between event mention pairs. Finally, a multi-layer perceptron captures these similarities to resolve coreferent events. Experimental results on the ECB+ dataset show that our proposed model outperforms several baselines and achieves competitive performance with state-of-the-art models.

Introduction

In Natural Language Processing, Cross-Document Event Coreference Resolution (CD-ECR) is essential for synthesizing information across vast corpora. The goal is to determine if event mentions in disparate documents—such as "the arrest" in a local news report and "the detainment" in a national broadcast—refer to the same real-world occurrence.

Historically, models have been "myopic," focusing predominantly on local context—the immediate sentence surrounding an event trigger. While effective for simple cases, this approach struggles with global document context and long-distance interactions. It fails to leverage the logical flow of a narrative, which often provides critical clues for coreference. This research proposes that the underlying rhetorical structure of a document—how ideas are linked through logic like "Cause," "Contrast," or "Elaboration"—serves as the necessary "connective tissue" to link these mentions accurately.

Methodology: Integrating Global Discourse Structure

The proposed framework addresses the limitations of local-only models by combining standard local embeddings with global discourse features. The architecture splits the information representation into two parallel paths: Local and Global.

The overall architecture of the proposed CD-ECR model, showing the parallel processing of local and global information.

Discourse Tree Construction and Feature Extraction

The core innovation lies in the Discourse Rhetorical Structure (DRS) Constructor. To handle cross-document pairs, the authors developed the ES-comp strategy. Since standard Rhetorical Structure Theory (RST) parsers are designed for single documents, the model uses a T5 model to compress the surrounding context of two documents into a single "pseudo-document." This allows the parser to build a unified tree representing the relationship between two mentions across document boundaries.

From this constructed tree, two primary global features are extracted:

  1. Lowest Common Parent (LCP): The specific node where the branches of the two event mentions meet, representing their most immediate shared rhetorical context.
  2. Shortest Dependency Path (DT-SDP): The sequence of rhetorical relations (e.g., NS-Elaboration → NS-Cause) connecting the two mentions. This path is encoded using a Bi-LSTM to capture the structural logic between events.
Detailed view of the discourse tree construction process (a) and the extraction of LCP and DT-SDP features (b).

These global features are concatenated with local representations (derived from RoBERTa or Longformer) and fed into a Multi-Layer Perceptron (MLP) to calculate a final pairwise coreference score.

Experimental Results and Insights

The model was evaluated on the ECB+ (Event Coreference Bank Plus) dataset, the current benchmark for CD-ECR. The results demonstrate that structural context significantly enhances performance.

  • State-of-the-Art Performance: Using Longformer as the base encoder, the model achieved a CoNLL score of 86.4, surpassing the previous state-of-the-art of 85.7.
  • The Impact of Discourse: Ablation studies revealed that adding discourse information to a RoBERTa baseline boosted the CoNLL score from 81.1 to 85.7. This substantial jump validates the hypothesis that rhetorical relations are often more stable signals for coreference than vocabulary alone.
  • Feature Importance: The Shortest Dependency Path (DT-SDP) emerged as the most influential global feature. It provides a structural map of how two events are logically related, which is particularly useful for handling "singleton" events (events mentioned only once) where local context might be sparse.

Conclusion

By treating discourse structure as a semantic bridge, this research moves beyond the "myopic" limitations of previous CD-ECR models. The success of the ES-comp strategy and the DT-SDP feature highlights that how a story is structured is just as important as the specific words used to tell it. While the computational overhead of discourse parsing remains a challenge, the significant gains in accuracy suggest that global structural information is a vital frontier for event coreference resolution. Future work may explore optimizing these parsers or extending the approach to cross-modal environments.

Summary updated

Abstract

Cross-document event coreference resolution (CD-ECR) is a task of clustering event mentions across multiple documents that refer to the same real-world events. Previous studies usually model the CD-ECR task as a pairwise similarity comparison problem by using different event mention features, and consider the highly similar event mention pairs in the same cluster as coreferent. In general, most of them only consider the local context of event mentions and ignore their implicit global information, thus failing to capture the interactions of long-distance event mentions. To address the above issue, we regard discourse structure as global information to further improve CD-ECR. First, we use a discourse rhetorical structure constructor to construct tree structures to represent documents. Then, we obtain shortest dependency paths from the tree structures to represent interactions between event mention pairs. Finally, we feed the above information to a multi-layer perceptron to capture the similarities of event mention pairs for resolving coreferent events. Experimental results on the ECB+ dataset show that our proposed model outperforms several baselines and achieves the competitive performance with the start-of-the-art baselines.
BibTeX
@inproceedings{chen2023cross,
  author = {Xinyu Chen and Sheng Xu and Peifeng Li and Qiaoming Zhu},
  title = {{Cross-Document Event Coreference Resolution on Discourse Structure}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2023},
  url = {https://openreview.net/forum?id=f91Eeyzjtp},
}
Figure