A graph-based method for Chinese address matching

Address matching is a substantial task in location-based services. Currently, major address matching methods either perform rather badly on unstructured data or fail to extract adequate semantic information of address elements. In this article, we propose a graph-based method that can deal with both sides of the problem. First, we use a pretrained transformer neural network to handle address tokenization. Then we parse address tokens into address elements according to their parts of speech. Then the node2vec and tf-idf technique is used to generate node embeddings for each address element. Finally, an address matching graph convolutional network is applied to do the address matching work. We have carried out a series of experiments on a real-world Chinese address corpus, to further evaluate the impacts of our methods. The experimental results indicate our method achieves higher scores than the state-of-the-art methods.