Graph embedding, which aims to project a graph into a low-dimensional space, has received significant attention. Semi-supervised graph embedding takes advantage of labeled data, showing promising performance. However, real-world datasets are frequently completely-imbalanced (i.e., zero-shot), wherein certain node types have no labeled instances. This scenario poses a formidable challenge for conventional graph embedding models, resulting in suboptimal performance across various downstream tasks. In this paper, we present a novel heterogeneous graph embedding method, termed Semantic Heterogeneous Graph Convolutional Network (SHGCN). Our method integrates a semantic module into the heterogeneous graph convolutional network, facilitating the extraction of shared features from nodes of identical types within the original graph. Through the extraction of semantic features, SHGCN maps nodes of the same type in the original graph to a low-dimensional latent vector space, promoting proximity and enhancing the performance of the model in zero-shot scenarios. Extensive experiments on three real-world heterogeneous graph datasets demonstrate the effectiveness of the proposed method.