ZERO-SHOT COMPOSITIONAL EVENT DETECTION VIA GRAPH MODULAR NETWORK
MetadataShow full item record
Humans are known to have the capability of understanding events by composing different atomic concepts, even for event types that have never been seen before. However, event detection has been so far treated as a sequence tagging problem in literature. Despite the increasing accuracy obtained on benchmarks such as ACE, current supervised sequence tagging models lack the compositional generalization ability. We present a model that is able to achieve zero-shot compositional generalization for event detection. Our model, named compositional graph modular network (CGMN), proposes two separate graph neural networks to obtain compositional semantic representations for sentences and events respectively. Meanwhile, it ties graph-based event representations with the weight parameters of an event matching layer, so that the semantic representations for sentences and events can be connected with each other, thereby achieving zero-shot recognition of new events using only their constituent atomic concepts. Our experiments on the ACE 2005 dataset as well as our collected Twitter event dataset show that, CGMN significantly outperforms state-of-the-art event detection methods on unseen classes and demonstrate strong zero-shot compositional generalization capabilities.