A Spatially Explicit Pretrained Language Model for Named Geographic Entity Recognition and Localization

Keywords: Text Mining, Named Entity Recognition, Geographic Entity Localization, Spatially Explicit AI
Abstract Type: Paper Abstract

Abstract

Named geographic entities (or geo-entities), the fundamental units of many geospatial datasets, are ubiquitous in unstructured data. While there has been an abundance of studies taking advantage of pretrained language models to recognize and extract such geo-entities from text, most, if not all, of them fail to explicitly consider the geospatial attributes, including geographic coordinates and contexts. We hypothesize that incorporating such knowledge shall augment the performance of this information retrieval task and thereby propose a novel spatially explicit language model. Built upon BERT backbone, our model takes in geospatial embeddings of the entities and their contexts in a manner of multimodal learning. Furthermore, the extracted geo-entities are localized on maps accordingly, which is a nontrivial follow-up task considering the heterogeneous nature of geo-entities (being points, lines, or polygons). This text and map interacted framework not only enhances the recognition of named geographic entities in enormous text data, but facilitates the understanding of text content from a geospatially explicit perspective.