Novel Robust Span-based Model for Event Detection

Abstract

Event detection (ED) aims to identify events and their types from the given text, which is the vital subtask of event extraction. However, poor quality training corpus can seriously affect the validity of current dominating deep learning-based techniques. In order to improve the robustness of the model, this paper converts the ED task into the named entity recognition (NER) task, and uses the span-based NER framework to identify and classify event triggers. In addition, we ameliorate the framework through data cleaning strategy, hierarchical learning rate strategy, sentence-level loss function and model ensemble strategy. The experimental results suggest that our approach achieves state-of-the-art performance on the widely used Chinese dataset. Moreover, the deep analysis has confirmed the effectiveness of the improvement strategy in resisting data pollution.