Abstract
Event Coreference Resolution (ECR) is the task of identifying all mentions of a real-world event across one or more texts. Traditionally, this is handled via a linear pipeline: resolving within-document (WD) links first, followed by cross-document (CD) links. This paper introduces an iterative approach that alternates between WD and CD resolution, leveraging "second-order" relationships between event chains to achieve state-of-the-art results on the ECB+ corpus.
Introduction
Identifying coreferent event mentions is notoriously difficult because the same event can be described using highly varied lexical forms (e.g., "the match," "the victory") or appear in contexts where key details like participants or locations are omitted.
Standard approaches often treat WD and CD coreference as identical tasks or follow a strict linear pipeline. However, the authors argue this is suboptimal. WD and CD links have distinct "signatures"—CD links typically require stricter evidence and more environmental context. Furthermore, information gained from a CD merge (such as a location mentioned in one document but not another) can provide the necessary evidence to resolve a previously ambiguous WD link.
Methodology: Iterative Merging and Second-Order Relations
The proposed model employs a two-stage process that breaks the traditional linear pipeline in favor of a virtuous cycle of information propagation.
1. Iterative WD and CD Merging
The system uses two distinct neural-network-based pairwise classifiers. The WD Classifier focuses on event lemmas, parts-of-speech, and semantic arguments (who, what). The CD Classifier incorporates an LSTM layer to capture broader sentential context, reflecting the need for more "environmental" clues in cross-document linking.
In each iteration, the system performs WD merges followed by CD merges. After merging, it propagates arguments (merging the "who/where/when" info) into the new clusters. This enriched information serves as the input for the next iteration, enabling the discovery of more difficult links.
2. Exploiting Second-Order Inter-dependencies
Once iterative merging concludes, the system analyzes how different event clusters relate to one another. If two event mentions share the same dependency relationship with a third, already-coreferent event, they are likely coreferent themselves. For instance, if "the shooting" and "the attack" both have a dependency relationship with a specific "trial" event, the model can infer they refer to the same incident even if their surface forms differ significantly.
Experimental Setup and Results
The model was evaluated on the ECB+ (EventCorefBank+) corpus, the gold standard for ECR, which includes topic-relevant documents designed to challenge models with similar but distinct events.
Key Insights: WD vs. CD Signatures
A critical finding of the study is that WD and CD tasks rely on different features. As shown in the feature weight analysis, the CD classifier relies heavily on the cosine similarity of context embeddings, whereas the WD classifier is more dependent on specific event words and their direct arguments.
Performance Metrics
The model outperformed previous state-of-the-art methods (such as the HDDCRP model) across all standard metrics:
- Cross-Document: Achieved a CoNLL F1 score of 63.6 (a 4.9% absolute improvement).
- Within-Document: Achieved a CoNLL F1 score of 68.9 (a 2.1% improvement).
The iterative process proved highly effective; while most merges occur in the first pass, subsequent iterations successfully capture nuanced, synonymous links like (take over, replace) and (win, victory).
Conclusion and Future Work
By treating WD and CD as mutually beneficial tasks rather than a linear sequence, this model effectively propagates information across document boundaries. However, challenges remain regarding "light verbs" (e.g., "said," "gave") which appear in similar contexts but refer to distinct instances. Future work will explore incorporating discourse-level features, such as the physical distance between mentions, to further refine resolution in argument-sparse contexts.
Summary updated