Event Coreference Resolution by Iteratively Unfolding Inter-dependencies among Events

Prafulla Kumar Choubey, Ruihong Huang

Abstract

We introduce a novel iterative approach for event coreference resolution that gradually builds event clusters by exploiting inter-dependencies among event mentions within the same chain as well as across event chains. Among event mentions in the same chain, we distinguish within- and cross-document event coreference links by using two distinct pairwise classifiers, trained separately to capture differences in feature distributions of within- and cross-document event clusters. Our event coreference approach alternates between WD and CD clustering and combines arguments from both event clusters after every merge, continuing till no more merge can be made. And then it performs further merging between event chains that are both closely related to a set of other chains of events. Experiments on the ECB+ corpus show that our model outperforms state-of-the-art methods in joint task of WD and CD event coreference resolution.

Introduction

In Natural Language Processing (NLP), event coreference resolution is the task of identifying all mentions of a specific real-world event (e.g., "the shooting," "the attack," "the tragedy") and clustering them together. This task is notoriously difficult because event mentions can vary significantly in surface form (e.g., "killed" vs. "murder") and often omit crucial details like participants or timestamps.

Researchers typically distinguish between Within-Document (WD) coreference and Cross-Document (CD) coreference. Historically, models either treated them as identical tasks or followed a rigid two-step process: grouping mentions within a document first, then linking those groups across documents. This paper argues that such approaches are insufficient because WD and CD links have distinct feature "signatures," and early errors in a sequential pipeline can propagate through the system.

Methods

The authors propose a two-stage iterative framework designed to build event clusters gradually.

Stage 1: Iterative WD and CD Merging

Instead of a one-pass system, the model alternates between WD and CD merging. After every merge, it performs argument propagation, combining the participants and locations from the merged mentions. This enriched information serves as additional evidence for subsequent iterations.

  • Distinct Classifiers: The system utilizes two separate neural network classifiers. The WD classifier focuses primarily on event lemmas and arguments. The CD classifier incorporates a Long Short-Term Memory (LSTM) layer to capture the broader context surrounding a mention, recognizing that cross-document links require more stringent contextual evidence.

Stage 2: Second-Order Inter-dependencies

Once initial clusters stabilize, the model identifies "second-order" links. If two event mentions (A and B) are both related to a third, already-known coreferent event (C) in the same way—for instance, sharing the same dependency relation to a specific "trial" event—the model considers A and B likely to be coreferent as well.

The iterative event coreference resolution process showing WD, CD, and second-order merges.

Experimental Setup

The model was evaluated on the ECB+ (EventCorefBank+) corpus, the gold standard for this task, containing nearly 1,000 documents and over 7,000 event mentions. The authors followed the standard split: topics 1-20 for training, 21-23 for validation, and 24-43 for testing.

Key Insights and Results

Feature Distribution Analysis

A key insight of this work is that WD and CD coreference are fundamentally different. Through a feature weight analysis, the authors demonstrated that WD coreference relies heavily on word similarity (lemmas), whereas CD coreference depends much more on surrounding context and semantic roles.

Feature weight analysis comparing Within-Document (WD) and Cross-Document (CD) coreference.

Performance

The iterative model significantly outperformed previous state-of-the-art methods:

  • Superior Accuracy: On the CoNLL F1 metric, the full model achieved 63.6 for Cross-Document and 68.9 for Within-Document resolution.
  • Comparative Gains: It outperformed the previous leading model (HDDCRP) by 4.9% on CD and 2.1% on WD tasks.
  • Ablation Success: Results showed that adding "2nd Order Relations" consistently improved performance, proving that modeling how events relate to other events provides a powerful signal for coreference.

Conclusion

While the iterative approach provides substantial gains, challenges remain. The system still struggles with sparse arguments, where mentions have very little local context, and reporting verbs (e.g., "said," "told"), which often share similar contexts but refer to distinct speech acts. Future work will explore discourse-level features, such as physical distance between mentions, to further refine clustering accuracy.

Summary updated

Abstract

We introduce a novel iterative approach for event coreference resolution that gradually builds event clusters by exploiting inter-dependencies among event mentions within the same chain as well as across event chains. Among event mentions in the same chain, we distinguish within- and cross-document event coreference links by using two distinct pairwise classifiers, trained separately to capture differences in feature distributions of within- and cross-document event clusters. Our event coreference approach alternates between WD and CD clustering and combines arguments from both event clusters after every merge, continuing till no more merge can be made. And then it performs further merging between event chains that are both closely related to a set of other chains of events. Experiments on the ECB+ corpus show that our model outperforms state-of-the-art methods in joint task of WD and CD event coreference resolution.
BibTeX
@inproceedings{choubey2017event,
  author = {Prafulla Kumar Choubey and Ruihong Huang},
  title = {{Event Coreference Resolution by Iteratively Unfolding Inter-dependencies among Events}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2017},
  url = {https://openreview.net/forum?id=Wmp3KjVkK3U},
}
Figure