Abstract:In order to solve the address matching problem between power address database and external address database and ensure the accuracy of power address, and realize the data information sharing and interworking between the power system and the external system, an address matching model based on "Retriever-Discriminator" architecture is proposed. Firstly, the detailed structure of the address matching model is introduced, including address retriever to narrow the search scope and the address discriminator to distinguish whether the address matches eventually. The Address Retriever is constructed based on the word frequency-inverse document frequency algorithm. The Address Discriminator is constructed based on the Chinese pre-training language model NEZHA. A negative sample training method is proposed to improve the discrimination effect of the model. Two datasets used in the experiment are introduced in detail. The experimental results show that the power address matching model based on “Retriever Discriminator” architecture can accurately find the address matching with the power address from the external address database. The Address Discriminator can find the matching address from multiple candidate address pairs with F1 score more than 0.99.