Assigning semantically relevant, real-world locations to documents opens new possibilities to perform geographic information retrieval. We propose a novel approach to automatically determine the latitude-longitude coordinates of appropriate Wikipedia articles with high accuracy, leveraging both text and metadata in the corpus. By examining articles whose base-truth coordinates are known, we show that our method attains a substantial improvement over state of the art works. We subsequently demonstrate how our approach could yield two benefits: (1) detecting significant geolocation errors in Wikipedia; and (2) proposing approximated coordinates for hundreds of thousands of articles which are not traditionally considered to be locations (such as events, ideas or people), opening new possibilities for conceptual geographic retrievals over Wikipedia.

We study the problem of deriving geolocations for Wikipedia pages. To this end, we introduce a general four-step process to location derivation, and consider different instantiations of this process, leveraging both textual and categorical data. Extensive experimentation shows that our methods provide good precision-recall trade-offs and improvements over text-only methods. Hence, our system can be used to augment the geographic information of Wikipedia, and to enable more effective geographic information retrieval.