Automatic extraction of destinations, origins and route parts from human generated route directions

Researchers from the cognitive and spatial sciences are studying text descriptions of movement patterns in order to examine how humans communicate and understand spatial information. In particular, route directions offer a rich source of information on how cognitive systems conceptualize movement patterns by segmenting them into meaningful parts. Route directions are composed using a plethora of cognitive spatial organization principles: changing levels of granularity, hierarchical organization, incorporation of cognitively and perceptually salient elements, and so forth. Identifying such information in text documents automatically is crucial for enabling machine-understanding of human spatial language. The benefits are: a) creating opportunities for large-scale studies of human linguistic behavior; b) extracting and georeferencing salient entities (landmarks) that are used by human route direction providers; c) developing methods to translate route directions to sketches and maps; and d) enabling queries on large corpora of crawled/analyzed movement data. In this paper, we introduce our approach and implementations that bring us closer to the goal of automatically processing linguistic route directions. We report on research directed at one part of the larger problem, that is, extracting the three most critical parts of route directions and movement patterns in general: origin, destination, and route parts. We use machine-learning based algorithms to extract these parts of routes, including, for example, destination names and types. We prove the effectiveness of our approach in several experiments using hand-tagged corpora.