A Neural Network Application To Database Quality: The Address Normalization Problem

In this ,paper, we investigate the problem of normalization of mailing addresses. Our approach is a "hybrid" approach. The two ingredients in our approach are the address dictionary and the scoring system. To start, our system is taught to recognize a set of addresses, the training set. Thus training process leads to the construction and update of the address dictionary which meanings of various grouped components of an address, which is called tokens. We assign weights to different meanings based on the frequencies of use. to normalized a new address, our system tries first to recognize the various components by looking up the address dictionary, and when multiple meanings are encountered, a scoring function is used to resolve the ambiguity.