With the development of online social networks such as Weibo, WeChat and Facebook, network representation learning has drawn widespread research interests from academia and industry. Traditional network embedding models exploit the spectral properties of matrix representations of graphs, which suffer from both computation and performance bottlenecks when applied to real world networks. Recently, a lot of neural network based embedding models are presented in the literature. They are computationally efficient and preserve the network structure information well. The vertices in the network are connected to various types of relations, which convey rich information. However, such important information are neglected by all existing models. This paper proposes NEES, an unsupervised network embedding model to encode the relations. It first obtains the edge vectors by edge sampling to reflect the relation types of the edges. Then, it uses the edge vectors to learn a low dimension representation for each node in the graph. Extensive experiments are conducted on several social networks and one citation network. The results show that NEES model outperforms the state-of-the-art methods in multi-label classification and link prediction tasks. NEES is also scalable to large-scale networks in the real world.