Automatic short answer grading (ASAG) is a challenging task that aims to predict a score for a given student response. Previous works on ASAG mainly use nonneural or neural methods. However, the former depends on handcrafted features and is limited by its inflexibility and high cost, and the latter ignores global word cooccurrence in a corpus and global interaction among the samples in datasets. However, ASAG requires this global information to learn the different expressions conveying the same meaning and the relations between the expressions and the grading labels. To address these limitations, we explore the use of a two-layer graph convolutional network (GCN) to encode the undirected heterogeneous graph of all student responses. The graph has sentence-level and word/bigram-level nodes. An edge is constructed between two nodes according to their inclusion or cooccurrence relationship. The sentence-level TF-IDF value or the PMI value is calculated as the edge weights to reflect the correlation degree between two nodes. Experimental results on the SemEval-2013 benchmark dataset and a two-subject dataset show that our model performs better.