Bi-directional feature learning-based approach for zero-shot event argument extraction

Abstract

Recent research has shown that event argument extraction (EAE) methods based on transfer learning and data augmentation emphasize the contribution of contextual features and labeled features to zero-shot EAE tasks, respectively. However, these methods suffer from knowledge transfer insufficiency and context generation bias challenges. In this paper, we propose a bi-directional feature learning-based approach for zero-shot event argument extraction (BiTer), which gains bi-directional transferable knowledge and mitigates context generation bias. Specifically, BiTer contains source and target model training. During source model training, BiTer co-trains the contextual and labeled feature learning tasks on the source dataset. This step enables the target model to acquire bi-directional transferable knowledge, providing more appropriate feature representations for target events. In target model training, BiTer leverages the large language model to produce pseudo-arguments, and then the knowledge-embedded model generates training data of the target events based on them. This step mitigates context generation bias and makes BiTer learn a more comprehensive and precise feature of the target event. Extensive experiments on RAMS, WIKIEVENTS and ACE2005 have demonstrated BiTer achieves a new state-of-the-art level, with F1 in the zero-shot setting outperforming the baseline model by 4.6%, 7.5% and 0.4%, respectively.