Multi-hierarchical error-aware contrastive learning for event argument extraction

Event argument extraction (EAE) aims to identify the spans and roles of arguments for the given event type. Deep learning-based EAE methods, especially generation-based methods, have achieved strong performance. However, constrained by supervised training with correct labels, these approaches struggle to discriminate potential extraction errors, manifesting as prediction omissions, redundancy, deviation, and boundary shifts, which limit the downstream applications of EAE. In this paper, we explore strategies for improving the representation learning capability of generative models to circumvent these potential errors. We reformulate the EAE as a template-filling task and propose ERCL, a multi-hierarchical error-aware contrastive learning framework. Specifically, we first design knowledge-free data augmentation algorithms, which generate negative templates covering known potential errors without introducing any external knowledge. Based on these negative samples, the model learns to identify the distribution of potential extraction errors from both sentence and span hierarchies under the guidance of contrastive learning, thus populating templates with correct argument spans. Extensive experiments on ACE05, RAMS and WIKIEVENTS demonstrate that ERCL outperforms other state-of-the-art methods on both sentence and document-level datasets. Furthermore, ERCL also shows strong performance in low-resource scenarios.