To accurately identify construction quality risk factors and analyze their underlying mechanisms, this study develops an integrated framework combining unsupervised semantic clustering with causal network analysis. The M3E-large sentence embeddings and latent Dirichlet allocation (LDA) model are applied for text vectorization and topic extraction, followed by keyword clustering using an improved canopy K-means algorithm. A systematic classification framework of risk factors is further established through three-level coding and a priori association rule mining. Subsequently, the interpretive structural modeling (ISM) method is employed to map factor dependencies and clarify hierarchical relationships, while a Bayesian network (BN) is constructed for risk assessment to quantify causal relationships and sensitivity. Based on 2,047 defect entries derived from 785 construction quality reports and five management manuals collected in China between 2014 and 2024, the results show that the proposed clustering model achieves better key information extraction performance than the traditional LDA (F1=0.81 versus 0.52). Compared with the unconstrained BN (75.3%), the ISM-BN demonstrates higher mean accuracy (83.7%) on the test set, indicating stronger stability. From real construction cases, 20 high-frequency factors, 25 key factors, and 30 sensitive factors are identified. Comprehensive analysis further summarizes five major categories of core driving risks: institutional system; structural safety; material equipment; technical compliance; and measurement monitoring. Among these, nine factors, including inadequate training system, undeveloped management system, supporting node failures, concrete structure deficiencies, substandard material processing and forming, mechanical transmission failures, coating surface treatment deficiencies, poor drainage slope and joint leakage, and excessive measurement errors, are identified as core drivers. The findings suggest that construction quality management should shift from reactive remediation to proactive prevention, emphasizing institutional development, personnel training, and process monitoring. This study provides an effective approach for extracting information from unstructured texts and analyzing complex risk mechanisms, thereby improving the efficiency of human error management.
Su et al. (Tue,) studied this question.