Better Exploiting BERT for Few-Shot Event Detection

Aboubacar Tuo, Romaric Besançon, Olivier Ferret, Julien Tourille

Abstract: Recent approaches for event detection rely on deep supervised learning, which requires large annotated corpora. Few-shot learning approaches, such as the meta-learning paradigm, can be used to address this issue. We focus in this paper on the use of prototypical networks with a BERT encoder for event detection. More specifically, we optimize the use of the information contained in the different layers of a pre-trained BERT model and show that simple strategies for combining BERT layers can outperform the current state-of-the-art for this task.