Abstract: The knowledge representation learning methods of spatial and semantic information fusion are studied by integrating geographic knowledge into spatial address. Spatial address data sets are trained on two classical translation models of TransE and TransH, and a comparative study is conducted through tuple classification and distance evaluation between vectors. The results of the study show that, in the representation learning task of address entities, the TransH algorithm model is significantly better than the TransE algorithm model in modeling complex relationships. The integration of semantic knowledge and spatial relationship can effectively solve the problems of address entities lacking correspondence between semantic similarity and spatial distance similarity. The fusion of the semantic relationship and the spatial relationship will be able to reveal more valuable information, help to complete the geographic knowledge graph and provide reference for the study of the geographic knowledge graph.