Learning Context-Aware Representation for Event Detection

Abstract

Current event detection(ED) task has confronted with low-resource scenarios which lead to the limited information extraction for event triggers. Regular methods shows weak performance of the appropriate trigger representation learning ability. In this paper, we integrate event trigger context similarity from the aspect of weakly-supervised learning for original event detection model. In model's training part, we use KL-divergence objective to optimize event trigger context representation. With the help of auxiliary objective of text similarity learning, the original trigger context could show more categoric semantic representation in feature space. Above all, with the architecture consisting of BERT and sequence labeling, we could detect event triggers with types simultaneously and as our experiment result shows on ACE2005 dataset, our approach shows advantages in event detection task.