Abstract
Event argument extraction (EAE) aims to extract arguments with given roles from texts, which have been widely studied in natural language processing. Most previous works have achieved good performance in specific EAE datasets with dedicated neural architectures. Whereas, these architectures are usually difficult to adapt to new datasets/scenarios with various annotation schemas or formats. Furthermore, they rely on large-scale labeled data for training, which is unavailable due to the high labelling cost in most cases. In this paper, we propose a multi-format transfer learning model with variational information bottleneck, which makes use of the information especially the common knowledge in existing datasets for EAE in new datasets. Specifically, we introduce a shared-specific prompt framework to learn both format-shared and format-specific knowledge from datasets with different formats. In order to further absorb the common knowledge for EAE and eliminate the irrelevant noise, we integrate variational information bottleneck into our architecture to refine the shared representation. We conduct extensive experiments on three benchmark datasets, and obtain new state-of-the-art performance on EAE.
Introduction
Event Argument Extraction (EAE) is a fundamental task in Information Extraction, focused on identifying the semantic roles (e.g., Attacker, Target, Instrument) associated with specific event triggers. Despite its importance, the field is hindered by two primary challenges: Format Inconsistency and the Noise Gap. Different datasets often utilize disparate schemas and labeling conventions, making it difficult to adapt models across domains. Furthermore, the high cost of manual annotation leads to a scarcity of labeled data for new event types.
To address these issues, we introduce UnifiedEAE, a transfer learning model designed to leverage knowledge from multiple datasets simultaneously. By distinguishing between format-specific nuances and universal event logic, UnifiedEAE achieves robust performance even in low-resource scenarios.
Methodology
Shared-Specific Prompt (SSP) Framework
UnifiedEAE is built upon the BART encoder-decoder architecture and utilizes a Shared-Specific Prompt (SSP) framework. This framework employs three distinct extractors: two are "format-specific," capturing the unique annotation styles of individual datasets, while the third is "format-shared," focusing on the underlying semantic logic common to all EAE tasks. Natural language prompts (e.g., "Person married Person at Place") are used to guide the extraction process, providing a universal interface that bridges different schemas.
Variational Information Bottleneck (VIB)
A critical innovation in UnifiedEAE is the integration of the Variational Information Bottleneck (VIB). The VIB acts as a mathematical filter for the shared representation. It encourages the model to compress information, retaining only the essential "signal" required for argument prediction while intentionally discarding dataset-specific "noise." This process ensures that the shared knowledge is truly universal and not biased by the quirks of a single source dataset.
Experiments and Results
Experimental Setup
We evaluated UnifiedEAE on three major benchmark datasets: ACE 2005 (sentence-level), RAMS (document-level), and WikiEvents (document-level). These datasets represent a wide variety of event types and semantic roles.
Performance and Low-Resource Gains
UnifiedEAE achieved new state-of-the-art (SOTA) results across all benchmarks. On the ACE 2005 dataset, it reached an Argument Classification (Arg-C) F1 score of 71.85, surpassing previous models like PAIE (69.80).
The model's most significant impact was observed in low-resource settings. As shown in the experiments, when provided with only 10 training samples from the target dataset, UnifiedEAE outperformed existing baselines by over 30 F1 points. This demonstrates the model's superior ability to transfer "event logic" from large source datasets to data-scarce target domains.
Conclusion
UnifiedEAE demonstrates that effective transfer learning for EAE requires both the acquisition of shared knowledge and the active filtering of dataset-specific noise. By combining the SSP framework with VIB, our model sets a new standard for performance and adaptability in event extraction. Future work will explore extending this architecture to full-pipeline event extraction, including trigger identification, and applying it to related tasks like Relation Extraction.
Summary updated