The significant advantages of point cloud data are presented in domains such as architectural reverse modeling, 3D reconstruction, and construction progress monitoring. Vast amounts of data are typically involved in the collection of point clouds for architectural structures, with the point clouds of components like beams and columns being particularly crucial. The challenges faced by current semantic segmentation methods for 3D point clouds when processing large-scale data include insufficient extraction of local features and suboptimal recognition accuracy. An enhanced approach for the semantic segmentation of large-scale point clouds of key architectural components using the RandLA-Net deep learning network was proposed. In this regard, the robustness of segmentation results was improved by incorporating a coordinate attention module in the local spatial encoding section. Furthermore, an extended channel attention module has been developed to strengthen the model’s capability in feature discernment, and a focal loss function has been introduced to effectively train the network, while addressing class imbalance issues within architectural point cloud scenes. Consequently, the efficient processing of architectural structure point cloud data and the extraction of key components are enabled. The performance comparisons and analyses conducted through experiments demonstrate that the original RandLA-Net model is outperformed by our model in terms of overall accuracy and component extraction precision in semantic segmentation of large-scale point clouds, thereby confirming the enhanced performance and practical value of the proposed method.