Using Event Semantics for Toponym Disambiguation

Introduction

Toponym disambiguation is the task of grounding ambiguous spatial locations in text (toponyms) by normalizing them to some structured representation (e.g., geo-coordinates, database entry, or location within a geographic ontology). This task proves to be quite difficult for some highly ambiguous locations. For example, there are over one thousand places named Springfield. To be able to assess the level of location name ambiguity, we have performed a two-tiered study. In the study, we used two gazetteers to count the number of locations with ambiguous names:

• •

  Geographic Names Information System (GNIS)¹ provided by the U.S. Geological Survey (USGS), which contains over two million locations and facilities within the United States.

• •

  GEOnet Names Server (GNS)² provided by the U.S. National Geospatial-Intelligence Agency (NGA), which contains over six million locations world-wide, excluding the United States.

We first determined that of more than two million unique location names in the two gazetteers, less than 20% are ambiguous. Then, to test the ambiguity of spatial locations in common use, we analyzed the SpatialML corpus (Mani et al., 2008), which contains 428 documents and 715 unique, manually-annotated location names. For these names, more than 80% are ambiguous. For example, our database contains 44 matches for “America,” including América/Mexico, America/Guinea-Bissau, and the United States of America; 24 matches for “Palestine,” including Palestine/Texas, and the Occupied Palestinian Territory; and 7 matches for “Baghdad,” including both the Iraqi city of Baghdad and the Iraqi Governate of Baghdad that contains the city. For more details on this case study, see Table 1.