Prior Knowledge Integrated with Self-attention for Event Detection

Recently, end-to-end models based on recurrent neural networks (RNN) have gained great success in event detection. However these methods cannot deal with long-distance dependency and internal structure information well. They are also hard to be controlled in process of learning since lacking of prior knowledge integration. In this paper, we present an effective framework for event detection which aims to address these problems. Our model based on self-attention can ignore the distance between any two words to obtain their relationship and leverage internal event argument information to improve event detection. In order to control the process of learning, we first collect keywords from corpus and then use a prior knowledge integration network to encode keywords to a prior knowledge representation. Experimental results demonstrate that our model has significant improvement of 3.9 F1 over the previous state-of-the-art on ACE 2005 dataset.