Abstract:

To address the difficulty of extracting critical hazard information from unstructured railway accident reports and to improve the intelligent management and control capabilities of railway transportation safety risks, a knowledge graph construction method for railway operational safety risk events was proposed. The method was based on a dictionary-enhanced Token Pair Linking (TPLinker) model. Typical train operation safety accident analysis reports from a railway bureau were used as the data source. First, a dictionary-enhanced Bidirectional Encoder Representations from Transformers (BERT) was introduced into the model encoding layer to strengthen domain-specific vocabulary representations, enabling the joint extraction of entities and relations from the textual corpus. Second, the Neo4j graph database was used to achieve the storage and visualization of the risk event knowledge graph. Finally, the Cypher query language was used, and knowledge question-answering technologies were explored to support complex intelligent question-answering and auxiliary decision-making scenarios. The results show that the improved TPLinker model achieves an F₁-score of 90.59% in the knowledge extraction task. Compared with the Copy Relation Representation Learning (CopyRRL) model, the Cascade Binary Tagging Framework (CasRel) model, and the standard TPLinker model, the knowledge extraction precision of the improved TPlinker model is improved by 20.79%, 4.46%, and 2.57%, respectively.

Key words: railway operational safety, risk events, knowledge graph, joint extraction, lexicon enhancement, Cypher query