Graph-structured neural networks build a graph over the sentences or entities in a document and use graph neural network layers to model cross-sentence dependencies for relation extraction. Propagating information across this graph structure gives the model memory and reasoning abilities that single-sentence models lack.