Aiming at the shortcomings of traditional graph neural network methods in point cloud semantic segmentation, such as high supervision accuracy requirement, node label transfer can only be unidirectional, and global information is not taken into account, this paper proposes a point cloud semantic segmentation method based on bidirectional attention mechanism. The algorithm segments the point cloud into superpoints and constructs a superpoint graph thus introducing the point cloud classification problem into the superpoint graph framework. After that, using the two-way attention module, it alternately pays attention to the superpoints and updates the superpoint features according to the weights of the neighboring superpoints, so as to achieve the two-way transfer of information. Also compared to previous graph pooling methods, this paper applies both maximum pooling and average pooling and combines the pooled features. In this paper, we use the public dataset Semantic3D for training and experiments, and the results show that the proposed method can effectively correct the annotation error and combine the long-range information, and the mIoU and oAcc of the dataset are 75.4% and 95.1%, respectively, which reflect better label delivery mechanism and higher classification accuracy compared with state of art methods.