Deep neural network based learning to rank for address standardization

Address standardization is the process of converting and mapping free-form addresses into a standard structured format. For many business cases, the addresses are entered into the information systems by end-users. They are often noisy, uncompleted, and in different formatted styles. In this paper, we propose a deep learning-based approach to the address standardization challenge. Our key idea is to leverage a Siamese neural network model to embed raw inputs and standardized addresses into a single latent multi-dimensional space. Thus, the corresponding of the raw input address is the one with the highest-ranking score. Our experiments demonstrate that our best model achieved 95.41% accuracy, which is 6.6% improvement from the current state of the art.