Extracting a Spatial Ontology from a Large Flickr Tag Dataset

We propose an easy framework for automatically constructing spatial ontologies that locate related concepts together in a space. The conventional graph representation is strong in showing direct relationships between entities, but it is difficult to process its topology when extracting features from the network, because similarity between networks is not well determined. Spatial ontologies are easy to cluster and classify according to the similarities or relationships between entities. We propose a method for creating a spatial ontology called "Associated Keyword Space" and apply it to 0.4M tag words collected from more than 1M images in Flickr. Tags in Flickr have many unknown word tags, but the spatial ontology can explain the clusters of meaning including unknown word tags. The results show that these unknown word tags can be found from neighbor tags that have clear meanings. As a result, an "area ontology" can be explained from the spatial ontology.