École Nationale Supérieure de
Techniques Avancées

Abstract

Within the context of Event Extraction, this research work addresses the trigger detec-tion task as a sequence labeling problem. We develop several approaches based on thepretrained BERT model. We start by a classical fine-tuning approach. Then, we exploremore advanced fine-tuning methods that highlight some important characteristics of thismodel. Furthermore, we investigate the impact of having different training and evaluationdata in terms of event types distribution and linguistic forms.