Network representation learning aims at learning low-dimensional representation for each vertex in a network, which plays an important role in network analysis. Conventional shallow models often achieve sub-optimal network representation results for non-linear network characteristics. Most network representation methods merely concentrate on structure but ignore text information related to each node. In the paper, we propose a novel semi-supervised deep model for network representation learning. We adopt a random surfing model to capture the global structure and incorporate text features of vertices based on the PV-DBOW model. The joint similarity between vertices achieved by combining network structure and text information is applied as the unsupervised component. While the first-order proximity in a network is used as the supervised component. By jointly optimizing them, our method can obtain reliable low-dimensional vector representations. The experiments on two real-world networks show that our method outperforms other baselines in the task of multi-class classification of vertices.