Geocoding location expressions in Twitter messages: A preference learning method

Abstract

Resolving location expressions in text to the correct physical location, also known as geocoding or grounding, is complicated by the fact that so many places around the world share the same name. Correct resolution is made even more difficult when there is little context to determine which place is intended, as in a 140-character Twitter message, or when location cues from different sources conflict, as may be the case among different metadata fields of a Twitter message. We used supervised machine learning to weigh the different fields of the Twitter message and the features of a world gazetteer to create a model that will prefer the correct gazetteer candidate to resolve the extracted expression. We evaluated our model using the F1 measure and compared it to similar algorithms. Our method achieved results higher than state-of-the-art competitors.