Constrained Multi-Task Learning for Event Coreference Resolution

Jing Lu, Vincent NgNAACL 2021 · DOI

Abstract

Event coreference resolution—determining if multiple mentions refer to the same real-world occurrence—is a complex task traditionally handled by error-prone pipelines. This paper presents a novel neural model that jointly trains six related information extraction tasks. By unifying entity and event coreference and enforcing soft consistency constraints through penalty functions, the model achieves state-of-the-art performance on the TAC KBP 2017 dataset, significantly outperforming previous joint and pipeline approaches.

Introduction

Event coreference resolution is inherently dependent on several auxiliary information extraction (IE) tasks, including trigger detection, argument extraction, entity coreference, and realis detection. Traditional pipeline systems suffer from error propagation, where inaccuracies in early stages (e.g., trigger identification) degrade final coreference performance. While joint learning offers a potential solution, scaling it across many tasks is computationally challenging and often leads to models that are too complex to converge effectively.

Methods

The authors propose a Constrained Multi-Task Learning (MTL) framework built upon a span-based neural model using SpanBERT-large. The architecture, illustrated in Figure 1, simultaneously learns six tasks: trigger detection, entity coreference, event coreference, anaphoricity (determining if a mention refers back to something), realis detection, and argument extraction.

Architecture of the Constrained Multi-Task Learning model showing the interaction between six NLP tasks and the application of soft constraints.

A key innovation is the unification of entity and event coreference into a single task, allowing for shared parameters and increased training data. Furthermore, the model incorporates "soft" consistency constraints. Unlike traditional "hard" constraints applied post-hoc, these are integrated into the learning process as penalty functions. This allows the model to penalize predictions that violate linguistic logic (e.g., coreferent events with conflicting subtypes) while maintaining flexibility in the face of noisy data.

Results

Evaluation on the TAC KBP 2017 Event Nugget Detection and Coreference dataset demonstrates the efficacy of the joint approach. The "Full Joint" model achieved an AVG-F score of 48.0 for event coreference, a substantial improvement over the previous state-of-the-art neural resolver (36.8) and non-neural resolver (37.9).

Performance comparison of the Full Joint model against previous state-of-the-art systems and pipeline baselines on the TAC KBP 2017 dataset.

As shown in Figure 2, the model also yields significant gains in trigger detection (64.5 F-score). Ablation studies confirmed that both the soft constraints and the unified coreference approach were critical to these gains, with their removal leading to performance drops of 1.0% and 0.8%, respectively. Explicitly training for anaphoricity also helped reduce false positives by filtering mentions that do not refer to previous entities or events.

Conclusion

The constrained multi-task learning framework represents a significant step toward a unified model for coreference resolution. While the model requires careful hyperparameter tuning and substantial computational resources, its ability to leverage inter-task dependencies through soft constraints sets a new benchmark for the field. Future work will explore extending this unified approach to a broader range of linguistic phenomena beyond entities and events.

Summary updated

BibTeX
@inproceedings{lu2021constrained,
  author = {Jing Lu and Vincent Ng},
  title = {{Constrained Multi-Task Learning for Event Coreference Resolution}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
  doi = {10.18653/v1/2021.naacl-main.356},
}
Figure