Research on RailBERT-based event extraction method for test cases of train control system on-board ATP

CHENG Ye; LI Kaicheng; WEI Guodong

doi:10.11860/j.issn.1673-0291.20230170

Abstract：In the laboratory testing of Automatic Train Protection(ATP)onboard equipment,the large volume and high complexity of test cases,combined with numerous specialized terms in the train con-trol domain,pose significant challenges for existing methods and models.These approaches often lack domain-specific knowledge,making it difficult to accurately interpret contextual information and auto-matically generate detailed structured representations.To address these challenges,this paper pro-poses an event extraction method for test cases based on the Rail Bidirectional Encoder Representa-tions from Transformers(RailBERT)model.First,a corpus of specialized terms in the train control domain is expanded and constructed using a neologism mining algorithm.A RailBERT model tailored to the train control system domain is then pre-trained with a Railway Whole Word Masking(RWWM)task to improve its understanding of domain-specific contexts.Then,an event extraction approach is developed to automatically extract the expected outcomes of onboard ATP test cases.The predefined event types and event theory elements are used to achieve comprehensive parsing and characterization of the expected results.Finally,the RailBERT is integrated with Bidirectional Long Short-Term Memory(BiLSTM)and Conditional Random Field(CRF)to enhance its ability to capture dependen-cies between sequence information and labels,thereby enabling more effective event extraction from test cases.The experimental results show that the proposed model achieves an F1 score of 90.3%on the test case event extraction dataset.This model accurately extracts predefined events from test cases and generates structured representations of the expected outcomes,providing a foundation for the implementation of automated testing.