Geographic information systems use databases to map keywords to locations. Currently, these databases are mostly created by a top-down approach based on geographic definitions. Problems are that (1) these databases only have information about addresses, location names, landmarks, and stores, and (2) if there are multiple candidate locations for a keyword, these databases do not have the information about which location is popular. A bottom-up approach which targets actual usage of keywords can address these problems. We propose a method to aggregate tagging data and extract locations related to a tag by using pairs of a tag and a geotagged resource. We use co-occurrence of a tag and a location and represent the locations related to a tag as a probability distribution over longitudes and latitudes. We apply our method to data on the photo sharing service Flickr. We experimentally confirm that our method can extract locations related to tags with high accuracy. Our bottom-up approach enables the extraction of location information that is unavailable using traditional top-down approaches.

View full abstract