Abstract: Map patterns in building groups have great importance in Cartographic Generalization and Multi-Scale Representation. On the basis of related research, a graph match approach is proposed to recognize the typical letter-like patterns in building groups. Typical letter-like pattern templates are extracted and analyzed, and selected as elementary units and described by a Attributed Relational Graph using attribute and structure parameters. A template library was established. Buildings to be abstracted and reduced are translated into Field Model based on the Attributed Relational Graph. Typical letter-like patterns are recognized by solving the imprecise sub-graph isomorphism problem with the Ullman algorithm. Experiments show that this approach is effective, feasible, and practical for typical letter-like pattern recognition and the results agree with human spatial cognition, providing a new concept in cartographic generalization.