With the increasing availability of laser point cloud data, research on how to extract rich point cloud feature information has become particularly important. Existing methods primarily focus on local feature learning while neglecting the relationship between point cloud positions and their features, and failing to model global information. To address this issue, an Adaptive Spatial Feature (ASF) module and the Multi-Scale Dilated Residual Block are proposed in this paper. The ASF consists of the Adaptive Feature Block and the Mixed Local Block, which can dynamically learn the relationship between point cloud positions and features, as well as eliminate uniform weighting. The Mixed Local Block combines local maximum feature data with local adaptive feature data to preserve local contextual details. The ASF is integrated into an encoder-decoder structure to form the ASF-Net network, and GroupFormer is introduced to extract global point cloud feature information. Experimental results demonstrate that ASF-Net exhibits outstanding semantic segmentation performance on the S3DIS and ScanNet v2 datasets, improving the accuracy of point cloud feature extraction.