Cross-document Coreference Resolution over Predicted Mentions

Arie Cattan, Alon Eirew, Gabriel Stanovsky, Mandar Joshi, Ido Dagan

Introduction

Coreference resolution—the task of identifying all mentions in a text that refer to the same real-world entity or event—is a cornerstone of natural language understanding. While within-document (WD) coreference has seen significant advancements through end-to-end deep learning models, cross-document (CD) coreference remains a formidable challenge. CD coreference requires identifying coreferring mentions across vast collections of documents, where no inherent linear order exists.

Historically, most CD coreference research has relied on "gold mentions"—pre-identified, human-annotated spans of text. This paper, presented at ACL-IJCNLP 2021, breaks this trend by introducing the first end-to-end model for CD coreference resolution that operates directly on raw text.

The Methodology

The authors adapt the well-known e2e-coref architecture for the cross-document setting. The pipeline involves four primary stages: span extraction using RoBERTa-Large, span pruning to manage computational complexity, pairwise scoring of candidate spans, and finally, agglomerative clustering.

The end-to-end model pipeline, from span extraction to agglomerative clustering.

As illustrated in Figure 1, this approach effectively decouples the task from the textual order of mentions. By using agglomerative clustering instead of traditional antecedent-based linking, the model can merge clusters of mentions across unrelated documents based on their pairwise similarity scores.

Experimental Results

The model was evaluated on the ECB+ (EventCorefBank Plus) dataset, which contains 982 documents organized into topics and subtopics. This structure is designed to test a model's ability to distinguish between lexically similar but distinct events. The authors evaluated their model in three scenarios: resolving events only, entities only, and a joint "ALL" setting.

Comparative performance of the proposed model against existing baselines on the ECB+ dataset.

As shown in Figure 4, the model achieves competitive results compared to state-of-the-art systems when using gold mentions, despite its simpler architecture. Crucially, the paper establishes the first baseline for the end-to-end task using predicted mentions. While there is a significant performance gap between gold and predicted mentions (e.g., 81.0 vs 54.4 F1 for events), this baseline provides a realistic starting point for future research.

Efficiency and Key Insights

A major contribution of this work is its efficiency. By encoding mentions once and using a simple Multi-Layer Perceptron (MLP) for pairwise scoring, the model is approximately 18 times faster than previous state-of-the-art systems, reducing inference time from 37 minutes to just 2 minutes.

The authors also highlight the "Singleton Effect." In WD coreference, mentions that appear only once are often ignored. However, in CD coreference, a singleton in one document may corefer with a mention in another. Accounting for these singletons is vital for accurate cross-document resolution and contributes to the performance difference seen when moving from gold to predicted mentions.

Conclusion

By removing the reliance on gold mentions and introducing an efficient, end-to-end clustering-based approach, this research sets a new standard for cross-document coreference resolution. While challenges remain—particularly regarding lexical ambiguity and temporal reasoning—this model provides a scalable and effective foundation for identifying entities and events across large-scale document collections.

Summary updated

BibTeX
@inproceedings{cattan2021cross,
  author = {Arie Cattan and Alon Eirew and Gabriel Stanovsky and Mandar Joshi and Ido Dagan},
  title = {{Cross-document Coreference Resolution over Predicted Mentions}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.findings-acl.453},
}
Figure