Dynamically modelling geospatial words in social media, in one aspect, generates a word set based on frequencies of words occurring in GPS annotated text data generated by a GPS-enabled device containing latitude and longitude coordinates. Locations are partitioned by mapping GPS coordinates in the GPS annotated text data to a set of discrete non-overlapped locations. A text stream contained in the GPS annotated text data is segmented into time windows. Footprints of locations in time windows are generated. Geospatial weights associated with words in the word set are generated based on localness of words determined based on the footprints. Words in a text message are extracted and scores are determined for the set of discrete non-overlapped locations associated with the words.