Geocoding Textual Documents Through a Hierarchy of Linear Classifiers

Most documents can be said to be related to some form of geographic context and, recently, geographic information retrieval has captured the attention of researchers from fields related to text mining and data retrieval, envisioning the support for tasks such as map-based document indexing, retrieval and visualization. In this paper, I empirically evaluate automated techniques, based on a hierarchical representation for the Earth’s surface and leveraging linear classifiers, for assigning geospatial coordinates of latitude and longitude to previously unseen documents, using only the raw text as input evidence. The obtained results were measured with models based on Support Vector Machines, over collections of geo-referenced Wikipedia articles in four different languages, namely English, German, Spanish and Portuguese. The best performing models were based on Support Vector Machines, obtaining state-of-the-art results corresponding to an average prediction error of 86 Kilometers, and a median error of just 8 Kilometers, in the case of the English Wikipedia collection. For the German, Spanish, and Portuguese collections, which are significantly smaller, the same method obtains an average prediction error of 62, 184 and 109 Kilometers, respectively, and a median prediction error of 5, 13, or 21 Kilometers.