A deep learning architecture for semantic address matching

Abstract

Address matching is a crucial step in geocoding, which plays an important role in urban planning and management. To date, the unprecedented development of location-based services has generated a large amount of unstructured address data. Traditional address matching methods mainly focus on the literal similarity of address records and are therefore not applicable to the unstructured address data. In this study, we introduce an address matching method based on deep learning to identify the semantic similarity between address records. First, we train the word2vec model to transform the address records into their corresponding vector representations. Next, we apply the enhanced sequential inference model (ESIM), a deep text-matching model, to make local and global inferences to determine if two addresses match. To evaluate the accuracy of the proposed method, we fine-tune the model with real-world address data from the Shenzhen Address Database and compare the outputs with those of several popular address matching methods. The results indicate that the proposed method achieves a higher matching accuracy for unstructured address records, with its precision, recall, and F1 score (i.e., the harmonic mean of precision and recall) reaching 0.97 on the test set.