Due to the short length of short text, there are problems such as data sparseness and semantic blurring in short text classification. This paper proposes a new graph convolutional network BTM_GCN, which uses the Biterm Topic Model（BTM） to train a fixed number of document-level potential topics on a short text dataset, and embeds it as a node in a text heterogeneous graph. Then in a heterogeneous graph, the document nodes are connected. Finally, the graph convolution network is used to capture the high-order neighborhood information between documents, words and topic nodes, thereby enriching the semantic information of the document nodes and alleviating the problem of short text semantic ambiguity. The experimental results on three English short text datasets show that the proposed method has better classification effect than the benchmark model.