Neural Embeddings for Populated Geonames Locations.

The application of neural embedding algorithms (based on architectures like skip-grams) to large knowledge bases like Wikipedia and the Google News Corpus has tremendously benefited multiple communities in applications as diverse as sentiment analysis, named entity recognition and text classification. In this paper, we present a similar resource for geospatial applications. We systematically construct a weighted network that spans all populated places in Geonames. Using a network embedding algorithm that was recently found to achieve excellent results and is based on the skip-gram model, we embed each populated place into a 100-dimensional vector space, in a similar vein as the GloVe embeddings released for Wikipedia. We demonstrate potential applications of this dataset resource, which we release under a public license.