Constructing Place Representations from Human-Generated Descriptions in Hebrew.

Space is a natural and indispensable part of the human communication form, mostly based on natural-spatial descriptions with varying lexical structures that rely on human spatial cognition and perception. This is a "geographic language" which machines do not understand, and accordingly do not properly process. Consequently, geographic information retrieval is limited due to the lack of rich and comprehensive textual-geographical databases required, for example, for spatio-query processes. While in English there exists a relatively rich set of libraries and tools, in Hebrew there is a void, with no automatic tools for addressing this problem. We propose a methodology that mimics human literal place descriptions, utilizing implicit geometries and topologies existing in geospatial databases. This study focuses on the first stage, which includes collecting a lingual dataset of human place descriptions with an online survey. Using Hebrew Natural Language Processes, place entities and their spatial relations were extracted from the survey descriptions. Similar place entities and relations were simultaneously extracted from OpenStreetMap database. Through place queries that rely on textual phrases from these two sources, human descriptions of places were geolocated. Finally, these locations were compared to retrieved locations acquired through Google maps API on survey descriptions - showing very promising results in accurately locating the described places.