Mining nearness relations from an n-grams Web corpus in geographical space

Abstract

Interacting with spatial data effectively requires systems that not only process references to locations, but understand spatial natural language. Empirical research has demonstrated that near is vague, asymmetric and context dependent. We explore near in language using Microsoft Web n-grams for expressions of the form A near*, where A are placenames referring to different spatial granularities, ranging from points of interest to large U.S. cities and * are autocomplete suggestions for placenames. Analyzing the extracted expressions requires consideration of semantic and referent ambiguity. With more than 200,000 expressions we show not only what is considered to be near at different scales, but also produce intuitive maps of nearness for different locations.