Substructures in graphs follow a power-law distribution as words in natural languages. Graphs have such connections to natural languages and thus can be modeled by the neural language models.