Abstract: Coal mine safety accidents are associated with a huge volume of data, but their utilization rate and structural degree are relatively low. The emergence of knowledge graph technology provides a novel approach to solving the problem of disordered and redundant data of this kind. Aiming at the insufficient mining of semantic information in the construction of knowledge graphs in the coal mine vertical domain, this study proposes a method for constructing a coal mine safety accident knowledge graph based on multi-granularity semantic fusion. First, textual data such as coal mine safety accident reports are collected, sorted out, and preprocessed. An ontology model for coal mine safety accidents is constructed based on the seven-step method, and data annotation is completed according to the established ontology. Second, the baseline model for entity extraction is optimized by fusing character and lexical features based on the self-attention mechanism. Meanwhile, a dependency matrix and a phrase structure matrix are introduced to improve the graph neural network architecture for the relation extraction task. Finally, experiments on entity and relation extraction based on multi-granularity semantic features are carried out to complete the construction of the coal mine safety accident knowledge graph. The experimental results show that the proposed entity-relation extraction model with multi-granularity semantic fusion achieves an F₁-score of 91.63% in coal mine safety accident entity extraction and 90.64% in relation extraction. Using natural language processing technology integrated with deep semantic information, this study successfully constructs a knowledge graph in the field of coal mine safety accidents, which effectively improves the performance of entity extraction and relation extraction models, and provides a feasible technical scheme and practical reference for the construction of vertical domain knowledge graphs.