Geospatial phrase grounding and disambiguation

Abstract

GeoCoder is a spatial reasoning system that converts natural language inputs into a set of precise spatial coordinates to display on a map. GeoCoder's spatial knowledge is represented in a set of ontologies. GeoCoder parses input phrases and adds location reference individuals to its ontology model. Relationships between location references are recognized based on mid-level structural patterns in the parsed phrase. GeoCoder grounds (or finds possible geometries for) location references in an iterative process, in which locations are grounded based on their relationships to previously grounded locations. GeoCoder improves upon previous systems by grounding and disambiguating at the phrase level, interpreting parses with rules that match mid level structure patterns, expressing disambiguation heuristics in ontologies, and improving scalability by separating grounding from reasoning about relationships.