Abstract: Point cloud semantic segmentation holds significant importance for architectural heritage documentation and preservation, with particular relevance to historic masonry structures such as ancient city walls. This study presents an innovative local attention-based approach to address several critical challenges in urban heritage point cloud analysis, including: (1) cross-domain adaptation, (2) severe class imbalance, (3) non-uniform point density distribution, and (4) unreliable RGB color information. The proposed framework introduces a novel neighborhood feature aggregation strategy that dynamically weights point contributions through learned attention scores, effectively modeling both intrinsic point characteristics and their contextual relationships within local neighborhoods. Our comprehensive evaluation demonstrates state-of-the-art performance on large-scale cultural heritage segmentation tasks, with quantitative results showing significant improvements in both segmentation accuracy and robustness to chromatic interference compared to existing methods. The method's superior performance is particularly evident in complex heritage scenarios where traditional approaches typically fail due to structural irregularities and material heterogeneity.