Transformer-Aware Graph Convolution Networks for Relation Extraction of Railway Safety Risk

The relation extraction of railway safety risk is important in constructing railway knowledge graph, which drives the development of intelligent transportation system through knowledge. However, existing models always concentrate on extracting relation from the universal text, fail to mine the railway text, especially the Chinese railway text. In this paper, we propose a novel relation extraction model named Transformer-aware Graph Convolution Networks (TGCN), which aims to comprehensively perceive railway semantics and improves the final results. Specifically, we construct a graph structure by the dependency tree to capture syntax features of input sentence, which can deep exploit the features contained in the sentence without introducing extra knowledge. In addition, we integrate these syntax features with contextual feature captured by the transformer mechanism. To better aggregate the railway semantic features, we use GCN to further encode the features confused by syntax and contextual features. Finally, a softmax classifier is used to identify the relation types. Extensive experiments on two real-world datasets show that, the proposed TGCN model achieves the state-of-the-art results. The TGCN provides a novel view to operate the railway safety risk management.