To do the text classification on the basis of VSM, and use the maximum common subgraph to measure two graphs� similarities are the relatively common methods, but these methods have not made full use of lots of semantic information spatial model contained, so the text classification performance is generally poor. In order to improve the classification results of the graph, on the basis of the structural equivalence, this paper further analyzes the maximum common substructure graph nodes and edges if it is a true semantic equivalence, and puts forward a kind of improvement text similarity metrics based on the graph space model. Then apply it to the text classification, the classification performance has been improved. Finally, verify the effectiveness of this method by experiment.