This paper investigates a graph-based semi-supervised learning algorithm, that is, label propagation algorithm for relation extraction. Labeled and unlabeled examples are represented as the nodes, and their distances as the weights of edges in the graph. The relation extraction tries to obtain a labeling function on this graph to satisfy the global consistency assumption. Experimental results on the ACE (automatic content extraction) corpus showed that this method achieves a better performance than SVM (support vector machine) when only very few labeled examples are available, and it also performs better than bootstrapping for the relation extraction task.