Nowadays, Graph Convolution Networks (GCNs) have flourished in the field of text classification, such as Text Graph Convolution Network (TextGCN). But good performance of those methods is based on building a graph whose nodes consist of an entire corpus, making their models transductive. Meanwhile rich label information has not been utilized in the graph structure. In this paper, we propose a new model named Inductive Light Graph Convolution Networks (ILGCN) with a new construction of graph. This approach uses labels and words to build the graph which removes the dependence between an individual text and entire corpus, and let ILGCN inductive. Besides, we simplify the model structure and only remain the neighborhood aggregation, which is the most important part of GCNs. Experiments on multiple benchmark show that our model outperforms existing state-of-the-art models on several text classification datasets.