Event detection from text using path-aware graph convolutional network

Event detection aims to detect events from text by locating event triggers and classifying them into predefined event types. Current state-of-the-art event detection methods benefit from integration of syntactic dependency into graph convolutional network (GCN). Despite the great success of GCN-based event detection methods, there are still two problems. Firstly, most GCN-based methods are designed as stacked structure to capture high-order contextual information, which will result in over-smoothing problem; secondly, dependency type information are not fully utilized in current GCN-based methods due to severe sparsity problem of some dependency types. In this paper, we propose P ath-Aware G raph C onvolutional N etwork (PGCN) model, shedding lights on simultaneously tackling these two problems. Specifically, PGCN is designed as flat structure to avoid over-smoothing problem, while path-aware aggregation is proposed to capture all-order contextual information and integrate dependency type information into feature space at the same time. Moreover, to deal with sparsity problem of some path types, we further adopt latent factor decomposition (LFD) technique by sharing parameters among different kinds of path. Our method is verified on the benchmark ACE 2005 English dataset. Experimental results show that our method gets 1.4% improvement on F1 score over state-of-the-art method, and performs more stably than other GCN-based methods with separate random initializations.