Joint Extraction of Hazard Source Knowledge in Integrated Utility Corridor Based on Knowledge Graph

The current knowledge graph-based joint extraction method for hazard source identification in integrated utility tunnels has been applied to emergency response management, assisting personnel in making rapid and accurate decisions. The core steps in building a knowledge graph involve entity recognition and relationship extraction. Joint extraction models overcome the limitations of pipeline extraction methods by enhancing the interaction between entities and relationships. However, existing joint models tend to overlook the semantic information between words and have room for improvement in the accuracy of extracting overlapping entities and relationships. This paper proposes an end-to-end, relation-specific attention network model, RAJE (Joint extraction model based on relational attention mechanism), to address semantic information and entity-relationship overlap issues. The RAJE model integrates active learning, data augmentation, BERT, a relation-specific attention network, and a sequence labeling approach to simultaneously extract all entities and relationships, demonstrating the model’s effectiveness in resolving nested entity and relationship problems. First, an active learning-based sampling method selects more valuable training data, and data augmentation is performed on the labeled dataset. Then, an end-to-end joint extraction method is used to obtain triplets. The joint extraction includes using the BERT model to obtain sentence representations, employing a relation-specific attention network to capture relation-based sentence features, followed by sequence labeling to extract entity pairs. Finally, the proposed model is evaluated using both a collected integrated utility tunnel fault engineering dataset and a public dataset.