Event causality identification network based on knowledge and syntactic structure
Event causality identification is an important task of relationship extraction, which has received much attention recent years. Most of the existing methods separate syntactic structure from the background knowledge information. The early causality extraction methods focus on the analysis of syntactic structure level. With the development of deep learning, the methods that use the pre-training model combined with background knowledge has become the mainstream. However, neither of the above two kinds of methods fully integrates the sentence information and external knowledge, resulting in different degrees of information loss. To address this problem, we proposed a novel model of event causality identification combining syntactic structure and background knowledge. Our model parses sentences into knowledge syntactic graph structures that contain both syntax and knowledge, and uses the graph convolution network for information fusion. It considers both syntax and knowledge information, which further enriches the event representation and performs effectively. In experiments on the widely-used dataset EventStoryLine, the F1 score of our model achieves 0.445, a 2.3％ improvement over existing methods.
event causality identificationsyntactic structuregraph convolution networknatural language processing
万方数据
维普
