Enrichment of a Pretrained Language Model for Extracting Topological Relations between Named Geographic Entities

Keywords: Relation extraction, Topological relations, Geospatial Deep Learning, Natural Language Processing
Abstract Type: Paper Abstract

Abstract

Named geographic entities (or geo-entities) are ubiquitous in unstructured data such as text. Extensive research has been conducted to recognize geo-entities from text using deep learning, yet limited work has addressed how to extract spatial relationships among them. Relation Extraction (RE), which focuses on detecting relations between entities mentioned in text data, is an important task in natural language processing, and plays a critical role in supporting many downstream applications, such as question answering and information retrieval. Existing spatial relations in RE studies are limited to simple location descriptions, such as "located in", without a deep understanding of topological relationships among geo-entities. Hence, to address this shortcoming, we have built upon the Bidirectional Encoder Representations from Transformers (BERT) backbone to propose a new spatially-aware language model which can capture richer and more accurate topological relationships. This model can improve the performance of topological relation extraction among named geo-entities.