Context-aware automated interpretation of elaborate natural language descriptions of location through learning from empirical data

Natural language descriptions of location can be complex, involving many different elements and often describing location by reference to other objects. Descriptions may be vague, and their meaning often depends upon the context within which the description has been expressed. Many previous approaches use mathematical models, focus on prepositions, and have had limited success and application. We present an approach to the interpretation of geospatial natural language expressions that uses a knowledge base of expressions for which human interpretations (in the form of degree of match to one of 50 geometric configurations) are known. Our approach interprets new expressions by finding the most similar knowledge base expression and adopting its meaning. We determine expression similarity using four different methods: element match; linguistic collocation approaches (Cosine); wordnet semantic network distance and a new approach that incorporates the contextual aspects of the expression including scale, geometry type, axial structure, image-schema and liquid/solid. As well as preposition, relatum and locatum, we consider spatial adjectives, adverbs, verb and sub-parts of the relatum and locatum. The method that incorporates context was the most successful of the four tested, selecting the same geometric configuration as human respondents in 69% of cases.