Metric Learning for Zero-Shot Event Extraction With Graph Convolutional Networks

Instead of deriving features from costly manual annotation, zero-shot approaches to event extraction maps every event mention to the predefined type in a target event ontology directly. Because of their natural advantage of representing structure features, there is a trend of applying graph convolutional networks to the zero-shot event extraction. However, previous works have focused on refining graph convolutional layers for better representation. In this paper, we explicitly frame the zero-shot event extraction as the combination of representation learning and metric learning. In addition, except the representation component, we explore various metric learning algorithms by utilizing different loss functions as well, which is easy to implement. Surprisingly, even with the vanilla graph convolutional network, replacing the loss function leads to the better performance over current state-of-the-art methods on two popular datasets, ACE2005 and GENIA.