Recently, three-dimensional point cloud semantic segmentation techniques based on sparse convolution have made great progress. However, sparse convolution causes a loss of global context information. In this study, a point cloud semantic segmentation method based on a sparse convolution and attention mechanism is proposed. Here, the attention mechanism is introduced into a sparse convolutional network to improve the network's ability to achieve global context information. However, extensive computation of the attention mechanism limits the applicability of the proposed method. Hence, to expand its usage while decreasing the amount of computation, spatial pyramid sampling is further introduced in the attention mechanism. Experimental results demonstrate that the proposed method achieves 71.825% of the average intersection over union (MIOU) on the Scannet V2 dataset and 70.5% on the S3DIS dataset, suggesting the proposed method's effectiveness and its superiority to the comparison method.