Matching Descriptions to Spatial Entities Using a Siamese Hierarchical Attention Network

Spatial entity descriptions are written in natural language based on the observations and understanding of spatial entities and often contain rich semantic information beyond GIS systems. Therefore, entity descriptions must be related to spatial entities in GIS systems. However, most previous studies of this issue were confined to place-type spatial entities, and other types of instance-level spatial entity matching have rarely been studied. In addition, existing matching methods require complex semantic analysis and manual feature engineering for the description text. In this paper, we focus on the matching of semantic similarity between spatial entities with rich text attributes and descriptions. We propose a semantic textual similarity matching model that incorporates a hierarchical recurrent structure with a focus on learning low-dimensional semantic vector representations of spatial entities and the corresponding descriptions. The model minimizes the distance between the vectors of matched pairs and maximizes the distance between the mismatched pairs of samples. The proposed siamese hierarchical attention network is trained and evaluated using a geological survey data set. The results show that the proposed model effectively captures the salient semantic information of spatial entities and the associated descriptions in the matching task and significantly outperforms previous state-of-the-art matching models.