Event Detection Model Based on Event Pattern and Type Bias

To address the problems of vague criteria for trigger word definition and the high cost of corpus annotation, a deep learning model for event detection called pattern and type based neural network (PTNN) is proposed. First, potential theorems are obtained based on entities' syntactic and semantic features. Then, the potential theorems are abstracted as roles. The embedding representation of PTNN is constructed by combining syntactic, semantic, and role features to enhance the representation of event patterns. Last, event detection and type determination are accomplished by using Bi-LSTM (bidirectional long short-term memory) with an event type-based attention mechanism. The model achieves event detection by enhancing event pattern features instead of identifying trigger words, thus avoiding the challenging problem of trigger word annotation. Such an approach demonstrates the positive effect of event patterns for event detection on neural networks. Experiments demonstrate that it improves the state-of-the-art of event detection by 3%.