A Novel Event Detection Model Based on Graph Convolutional Network

With the rapid development of society, economy, politics and science, there is a vast amount of collected daily news reports. How to detect news events and discover the underlying event evolution pattern has become an urgent problem. There have been many existing works to solve this problem, but most just use TF-IDF or LDA features to extract the limited semantic information, and the structural information of documents is also potential to be exploited. In this paper, we propose a novel Graph Convolutional Network based event detection model, named as NED-GCN, for news stream. The proposed model utilizes ConceptGraph to represent a document and fully takes semantic information and structural information of a document into account. Further, a Siamese Graph Convolutional Network (SiamGCN) is presented to calculate the similarity between document pair via shared weights for document embedding learning, and finally the learned document embeddings are clustered to generate events. Experimental evaluation on two real datasets shows that our method outperforms the state-of-art approaches in event detection.