Spatio-temporal provenance: Identifying location information from unstructured text

Spatio-temporal attributes represent two aspects of physical presence - space and time - which are integral to human activities. Space-time markers of an entity in conjunction with correlation with other networks such as movements in social network, the road/transportation network encodes a wealth of provenance information. With the advent of mobile computing and cheap and improved location estimation techniques, encoding such information has become commonplace. In this paper, we will focus on deriving such location provenance information from unstructured text generated by social media. As social media such as Facebook and Twitter are integrated with mobile devices, information generated by individuals in these networks gets tagged with spatial markers. We can classify such markers into explicit and implicit tags, where explicit tags encode the spatial data explicitly by providing the accurate location attributes. On the other hand, a lot of social network data may not encode such information explicitly. Our hypothesis in this paper is that the unstructured textual data contains implicit spatial markers at a fine granularity. We develop algorithms to support this hypothesis and evaluate these algorithms on data from FourSquare to show that the spatial category information can be identified with an accuracy of over 80%.