Chinese Address Standardization Based on seq2seq Model

Address information is important in many industries especially express delivery. However, since address is usually an unstructured text that has multiple expressions, the use of the address is confusing in the real world, for example, people usually simplify and express addresses according to their own habits. This leads to the challenge of missing and error of important information, which will reduce the efficiency for most of industries. In this paper, we propose an address standardization method based on the seq2seq model to solve the above challenges. We use the Gated Recurrent Unit (GRU) to learn the intrinsic link in the Chinese address and use attention mechanism to determine the weight of different parts of the address. To our knowledge, it is the first method that can complete the missing administrative address information and correct the error of address information without using additional information such as standard address database and geological element table. It is also the first time to use deep learning and attention mechanism in Chinese address standardization task. The experiment verifies that our method achieves high accuracy in the industry.