Enriching Graph Representations of Text: Application to Medical Text Classification

Graph based representations have been utilized to achieve state-of-the-art performance in text classification tasks. The same basic structure underlies knowledge graphs, large knowledge bases that contain rich information about the world. This paper capitalises on the graph of words model and enriches it with concepts from knowledge graphs, resulting in more powerful hybrid representations of a corpus. We focus on the domain of medical text classification and medical ontologies in order to test our proposed methods and analyze different alternatives in terms of text representation models and knowledge injection techniques. The method we present produces text representations that are both explainable and effective in improving the accuracy on the OHSUMED classification task, surpassing neural network architectures such as GraphStar and Text GCN.