Meta Learning for Event Argument Extraction via Domain-Specific Information Enhanced

Abstract

As a sub-task of event extraction, event argument extraction (EAE) aims at identifying arguments and classifying the roles they play in an event. Existing supervised learning based methods for EAE require large-scale labeled data which prevents the adaptation of the model to new event types. In this paper, we propose a meta-learning based EAE method, which aims at learning a good initial model that can adapt quickly to new event types with limited samples for training. For taking advantage of the consistency of event structures under the same domain, we introduce dynamic memory networks (DMN) to learn the domain-specific information. We conduct experiments under meta-learning setting to explore the scalability of our methods on EAE. The experimental results show that our method can learn general and transferable information that can be applied to the EAE of new event types which only contain few samples.