Abstract: To address the issues of significant subjectivity in manual analysis, low utilization of massive unstructured text date, and insufficient mining of dynamic coupling evolution mechanisms in coal mine safety risk identification, this study aims to utilize artificial intelligence text mining technology to achieve automated processing and deep knowledge discovery of accident big data, providing a scientific basis for precise risk prevention and control. Taking 1 709 investigation reports of underground coal mine accidents from 2015 to 2024 as the core corpus, a dynamic topic modeling and multidimensional analysis method based on Bidirectional Encoder Representations from Transformers based Topic Modeling (BERTopic) technology is proposed. First, pre-trained language models are utilized to transform accident texts into high-dimensional semantic vectors. Second, the Uniform Manifold Approximation and Projection (UMAP) algorithm is applied for nonlinear dimensionality reduction, combined with the Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) algorithm to adaptively extract 15 risk topics with significant semantic characteristics. Subsequently, integrating the Man-Machine-Environment-Management (MMEM) system theory, 94 key risk factors covering four dimensions are condensed from massive texts to construct a multidimensional risk knowledge system. Multidimensional analysis results indicate that risk topics present significant evolutionary laws in the time dimension. The intensity of management-related risks, such as hidden danger investigation and lack of supervision, continues to rise, indicating that soft management elements have become the bottleneck restricting the improvement of safety levels, while technical risks such as safety training and monitoring system failures are gradually alleviated. Differential analysis shows that the correlation intensity of electromechanical accidents with topics like “hidden danger investigation and on-site management” and “equipment maintenance” is significantly higher than that of other types, exposing weak links in the full life-cycle management of equipment. The study reveals the deep association and dynamic evolution mechanism of accident risk factors and verifies the effectiveness of the proposed method in processing unstructured safety text data. It provides data support and targeted guidance for constructing a dual prevention mechanism.