In this paper, we propose a framework for obtaining semantic labels of LiDAR point clouds and refining the classification results by combining a deep neural network with a graph-structured smoothing technique. In general, the goal of the semantic scene analysis is to assign a semantic label to each point in the point cloud. Although various related researches have been reported, due to the complexity of urban areas, the semantic labeling of point clouds in urban areas is still a challenging task. In this paper, we address the issues of how to effectively extract features from each point and its local surrounding and how to refine the initial soft labels by considering contextual information in the spatial domain. Specifically, we improve the effectiveness of classification of point cloud in two aspects. Firstly, instead of utilizing handcrafted features as input for classification and refinement, the local context of a point is embedded into deep dimensional space and classified via a deep neural network (PointNet++), and simultaneously soft labels are obtained as initial results for next refinement. Secondly, the initial label probability set is improved by taking the context both in the spatial domain into consideration by constructing a graph structure, and the final labels are optimized by a graph cuts algorithm. To evaluate the performance of our proposed framework, experiments are conducted on a mobile laser scanning (MLS) point cloud dataset. We demonstrate that our approach can achieve higher accuracy in comparison to several commonly-used state-of-the-art baselines. The overall accuracy of our proposed method on TUM dataset can reach 85.38% for labeling eight semantic classes.