Point cloud semantic segmentation refers to the accurate identification and classification of different objects in 3D scenes, providing high-precision environmental perception for applications such as autonomous driving and robot navigation. However, due to the unstructured features of large-scale 3D data and the occlusion problem, the local feature extraction ability is severely limited. In order to solve this problem, this paper proposes a novel point cloud semantic segmentation network TCLo. It mainly includes two modules: Local Feature Separator(LFS) and Local Feature Enhancer(LFE). Specifically, we propose LFS to construct the topology of the input point cloud to obtain richer local features, and then input the updated local features into the self-attention module to capture the global context information. The powerful and efficient local feature learning ability of LFS is combined with Transformer’s excellent global context modeling ability. LFE learns the down-sampled features more deeply, and identifies key points that contain richer structural information within subdivided local regions. The LFE prevents the loss of key points in local structures and provides strong support for subsequent up-sampling restoration tasks. To verify the effectiveness of the proposed modules, we design extensive ablation experiments on the S3DIS, ScanNetV2 and SemanticKITTI datasets, demonstrating the validity of the two modules.