With the reduction in the cost of acquiring 3D point cloud data and the improvement of GPU computing power, the demand for 3D point cloud semantic segmentation in numerous 3D visual scenarios, such as autonomous driving, industrial control, and MR/XR is growing, which further advances the development of deep learning methods in 3D point cloud semantic segmentation. Recently, many novel deep learning network architectures, such as RandLA-Net and Point Transformer, have been proposed. These new architectures have made breakthroughs in upgrading semantic segmentation accuracy with a lower computational load. However, the existing researches reviewing 3D point cloud semantic segmentation methods reported a lots of relatively early works, whose methods were gradually abandoned these years and cannot reflect the current research status without systematically organizing these new and efficient methods. In addition, the methods introduced in these researches were divided into categories depending on their different input data types, which cannot provide a progressive view of the relationship between methods using different network architectures and is not conducive to comparing the segmentation performance of different methods. Therefore, this paper is given with a focus on the mainstream of the last 3 years 3D semantic segmentation methods using different deep learning network architectures, and is conducted at 3 levels: First, two main 3D point cloud data acquisition methods with their usual datasets and the metrics to evaluate model performance are introduced. Second, a systematic review of 3D semantic segmentation methods based on different network architectures is organized, followed by a statistical analysis based on the performance of models, which are commonly used in different network architectures, on two 3D segmentation datasets S3DIS and ScanNet to these architectures, including their structure relevance, superiors and inferiors. Finally, an insightful discussion of the remaining challenges on both methodological and applications' viewpoints, and its leading potential research directions are given.  This paper provides a comprehensive review of recent 3-year research progress in 3D point cloud semantic segmentation. It summarizes different network architecture pipelines, illustrates their basic operations, compares the model performance for different architectures and discusses their highlights and shortcomings, more importantly, concludes the current challenges to provide promising research directions for the future. Thus, researchers can easily find the relevance and research hotspots between different 3D point cloud semantic segmentation methods based on these analyses. Through this paper, we aim to update the reviews on 3D point cloud semantic segmentation methods with a better viewpoint, highlight key properties and contributions of proposed methods so as to offer research directions from main challenges.