A domain semantics-enhanced relation extraction model for identifying the railway safety risk

The identification of railway safety risk is important in ensuring continuous and stable railway operations. Most works fail to consider the important relation between detected objects. In addition, poor domain semantics directly degrades the final performance due to difficulty in understanding railway text. To solve these challenging issues, we introduce the triple knowledge from knowledge graph to model the railway safety risk with the knowledge interconnection mode. Afterward, we recast the identification of railway safety risk as the relation extraction task, and propose a novel and effective D omain S emantics- E nhanced R elation E xtraction (DSERE) model. Specifically, we design a domain semantics-enhanced transformer mechanism that automatically enhances the railway semantics from a dedicated railway lexicon. We further introduce piece-wise convolution neural networks to explore the fine-grained features contained in the structure of triple knowledge. With the domain semantics and fine-grained features, our model can fully understand the domain text and thus improve the performance of relation classification. Finally, the DSERE model is used to identify the railway safety risk of south zone of China Railway, and achieves 81.84% AUC and 76.00% F1 scores on the real-world dataset showing the superiority of our proposed model.