Multi-view entity type overdependency reduction for event argument extraction

Event Argument Extraction (EAE) is a key component of event extraction, which has become a bottleneck that limits the overall performance of event extraction. As an entity-based extraction task, most EAE models focus on modeling complex interactions between entity mentions and event triggers. However, the strong correlation between entity types and argument role types has been overlooked in most EAE models, which disregard the possible negative effects of the correlation. In this paper, we study entity type dependency and conduct experiments to evaluate its effects on the overall performance for EAE. The experimental analysis shows that baseline EAE models suffer from varying degrees of entity type overdependency, which degrades the overall performance. To tackle this problem for EAE, we propose a novel multi-view entity type overdependency reduction model. The proposed model consists of two contrastive learning methods from different views and a cyclic training strategy. In particular, we propose a select-then-weigh contrastive learning method to achieve entity type overdependency reduction from the view of positive samples. And in parallel, we propose a pseudo-positive contrastive learning method to achieve entity type overdependency reduction from the view of negative samples. Moreover, the cyclic training strategy is designed to enable the two contrastive learning methods to collaborate efficiently. We have conducted experiments on the widely used ACE 2005 English dataset to evaluate the effectiveness of our proposed model. The experimental results show that our proposed model has outperformed the current state-of-the-art models for the EAE task.