Objective: Timely hazard detection during the construction of super high arch dams is crucial for reducing engineering accidents and ensuring project safety. Hazards in such settings are often hidden and diverse, making them difficult to detect during early-stage conventional inspections. However, fragmented hazard records at construction sites are crucial for identifying and detecting issues early, helping management personnel to promptly assess potential engineering risks. Methods: This study proposes an intelligent hazard identification method for super high arch dam construction using enhanced semisupervised contrastive learning. A multisource classification model for hazard text is developed to categorize and assess hazard types and levels from fragmented hazard texts, establishing a systematic hazard inspection framework. The model is built on the Transformer architecture, effectively capturing the semantic and positional relationships inherent in hazard descriptions. A contrastive learning module improves the Transformer by leveraging interclass relationships to amplify the differences between dissimilar samples. This significantly enhances classification accuracy, especially for multi-source attribute hazard categories. The method integrates self-supervised and supervised learning, emphasizing interclass distinctions while making use of label content. A memory bank mechanism decouples training batches, enabling comprehensive collection of negative samples, thereby enhancing the performance of semisupervised contrastive learning. Finally, the hazard category and level identification results are combined to visualize safety hazard distributions. Latent Dirichlet allocation (LDA) is used to extract latent clues for hazard risk inspection, constructing structured hazard inspection tables for different levels of risk. These tables allow managers to prioritize inspections in high-risk areas, enhancing the efficiency and precision of hazard detection. Results: The results show that the proposed classification model significantly improves hazard type and hazard level recognition tasks, with F1 score improvements of 4.9% and 3.3%, respectively. Multidimensional experiments were conducted to validate its significant advantages: 1) Analyzing the influence of different Memory Bank sizes on model performance highlighted the importance of batch decoupling batches and the selection of a robust number of negative samples; 2) Ablation experiments validated the contribution of each module to the model's performance improvement; 3) Dimensionality reduction clustering using t-SNE visually confirmed the contrastive learning module's ability to effectively group similar classification samples; 4) A comparison of infoNCE loss between this model and the base Transformer demonstrated the practical benefits of the contrastive learning module during training; 5) Performance comparisons with common classification models showed the proposed model's significant advantages in overall accuracy. The hazard category and level identification results are used to extract key topic information using the LDA topic model, revealing the potential risks present in the current hazard categories and levels. Taking "High-altitude fall" as an example, key topic clustering was applied to compile a complete hazard inspection clue table structured by hazard levels. Conclusions: The method enhances the precision and systematization of hazard identification during the construction of super high arch dams. It introduces a refined multi-source attribute hazard identification method, providing a novel approach to intelligent safety management in engineering and promoting the development of hazard management toward automation and intelligence.