Question answering remains a challenge for machines, partly due to the ambiguity of natural language and implied context, which often requires external knowledge to resolve. In this work, we present a general framework for incorporating such external knowledge (encoded as knowledge graphs) into question answering systems using graph convolutional neural networks. We applied our framework on top of Stochastic Answer Networks (SAN), a state-of-the-art method for question answering, and evaluated the system on the Stanford Question Answering Dataset 2.0. Our results show that leveraging knowledge brings significant improvements in terms of EM and F1 scores, validating the importance of incorporating external knowledge in understanding textual context.