Linking WHG: An analysis of SpaBERT's performance on WHG

Keywords: NLP, Spatial Resolution, Entity Linking, Historic Maps
Abstract Type: Paper Abstract

Abstract

In the recently published SpaBERT paper [Zekun. Et al. 2022] introduced a model that is able to characterize geo-entities based on their spatial context. The author's approach improves current baselines for geo-entity typing (if an entity is a hospital, gas station, school, etc.) and geo-entity linking (identifying the same geo-entity across different maps). They accomplish this through a combination of a standard BERT model for semantic understanding of pseudo-sentences and a spatial analysis component of the geographic locations of nearby entities for defining a spatial context. Our goal with this paper is to analyze the performance of SpaBERT on the geo-entity linking task between the WHG (World Historical Gazetteer) and Wikidata databases. Through this analysis, we hope to not only determine if SpaBERT performs well on this specific linking task but also provide insight into what factors play a role in the successes and failures of models like SpaBERT. The proposed analysis involves 1) analyzing the performance of SpaBERT as a whole on entity linking between WHG and Wikidata, 2) performing a comparative analysis between high-density map regions and low-density map regions shared by WHG and Wikidata, and 3) comparing performance on subsets of WHG and Wikidata that are semantically similar (i.e. Paris, Texas and Paris, France) and semantically dissimilar. Through this analysis, I hope to provide insight into how similar models might perform and provide recommendations for improving them moving forward.