Accuracy is critical when multiple databases are merged into a single system, because an error in a single record could lead to multiple mismatches. Address normalization is fairly common in database merging. We have developed a system to accurately and efficiently normalize mailing addresses. However, our system differs from other neural network architectures. Its key ingredients are an address dictionary and a scoring system. The scoring system is based on analog neural network systems, but the address dictionary follows a digital approach. The two key processes in our system are learning and address normalization. Learning is further split into dictionary creation updating and system parameters training.