A Multi-Format Transfer Learning Model for Event Argument Extraction via Variational Information Bottleneck

Jie Zhou, Qi Zhang, Liang He, Xuanjing Huang

Introduction

Event Argument Extraction (EAE) is a critical subtask of Information Extraction that involves identifying the specific roles played by entities within a triggered event—essentially determining "who did what to whom, where, and when." Despite its importance, EAE research is currently hindered by two primary obstacles: format inconsistency across disparate datasets and the high cost of manual annotation leading to data scarcity.

Existing datasets, such as ACE 2005, RAMS, and WikiEvents, often employ non-overlapping schemas. For instance, a "Conflict" event might be labeled with different granularities or role names (e.g., "Victim" vs. "Target") across different benchmarks. Simply merging these datasets often introduces noise that degrades model performance. To address these challenges, this paper introduces UnifiedEAE, a transfer learning framework designed to leverage multiple, inconsistently formatted datasets simultaneously.

Methodology: The UnifiedEAE Framework

UnifiedEAE addresses the problem of cross-format learning through two primary architectural innovations: the Shared-Specific Prompt (SSP) framework and the Shared Variational Information Bottleneck (SharedVIB).

Shared-Specific Prompt (SSP)

The SSP framework utilizes a multi-extractor architecture built upon the BART encoder-decoder model. Rather than employing a single monolithic extractor, UnifiedEAE uses three distinct extractors:

  1. Format-Specific Extractors: Two extractors are dedicated to capturing the unique nuances and annotation quirks of specific datasets.
  2. Format-Shared Extractor: A third extractor focuses on capturing the underlying semantic logic common to all event extraction tasks.

To provide a cohesive interface across different schemas, the model employs natural language prompts (e.g., "Person married Person at Place"). These prompts act as a universal translator, allowing the model to map disparate role names to a shared semantic space.

Shared Variational Information Bottleneck (SharedVIB)

The core challenge in multi-format learning is preventing dataset-specific "noise" from contaminating the shared knowledge base. UnifiedEAE implements the Variational Information Bottleneck (VIB) principle to act as a mathematical filter. The VIB objective encourages the "shared" extractor to compress its representations, retaining only the information that is most predictive of the task while discarding irrelevant, dataset-specific variations. This ensures that the shared representations capture the "essence" of the events rather than the idiosyncrasies of the labeling guidelines.

Experimental Evaluation

The authors evaluated UnifiedEAE against several major benchmark datasets:

  • ACE 2005: A classic sentence-level dataset with 33 event types.
  • RAMS: A document-level dataset featuring 139 event types and 65 roles.
  • WikiEvents: A document-level dataset with 50 event types and 59 roles.

Performance and State-of-the-Art Results

UnifiedEAE achieved new state-of-the-art (SOTA) results across all three datasets, outperforming strong baselines such as PAIE. Notably, on the ACE 2005 benchmark, the model achieved a 2.46% improvement in Argument Identification F1 score.

Low-Resource and Zero-Shot Capabilities

The framework's most significant advantage appears in data-constrained environments. In low-resource tests on WikiEvents using only 10 labeled examples, UnifiedEAE maintained an F1 score above 30, whereas previous SOTA models failed to produce meaningful results. Furthermore, the model demonstrated strong zero-shot potential, successfully extracting arguments for roles it had never encountered during training by leveraging the semantic similarities captured in the shared representation.

Discussion and Ablation Studies

Ablation experiments confirmed the necessity of the VIB component. Removing the VIB led to performance drops of up to 4 points, illustrating that "forgetting" dataset-specific noise is as vital to successful transfer learning as "learning" the shared features. The results suggest that the "Multiple" baseline—simply training on merged data—is often suboptimal due to conflicting labels, a problem UnifiedEAE effectively mitigates.

Conclusion

UnifiedEAE represents a significant step forward in making event extraction models more robust and data-efficient. By combining a shared-specific architecture with information bottlenecking, the model successfully navigates the complexities of multi-format learning. Future work may extend this approach to other NLP tasks such as Named Entity Recognition (NER) and Relation Extraction, and explore automated prompt generation to further scale the system to thousands of event types.

Summary updated

BibTeX
@inproceedings{zhou2022multi,
  author = {Jie Zhou and Qi Zhang and Qin Chen and Liang He and Xuanjing Huang},
  title = {{A Multi-Format Transfer Learning Model for Event Argument Extraction via Variational Information Bottleneck}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2022},
  url = {https://openreview.net/forum?id=0jrHHrq6uE},
}
Figure