Creating a corpus of geospatial language

The description of location using natural language is of interest for a number of research activities in geography, linguistics and cognitive science, including the development of methods for automated interpretation and generation of natural language to ease interaction with geographic information systems, as well as a number of related endeavours. For such research activities, examples of geospatial language are usually collected from the personal knowledge of researchers, or in small scale collection activities specific to the project concerned. This paper describes the process used to develop a more generic corpus of geospatial language. While the motivation for development was the authors’ ongoing research into natural language geospatial querying, it also has wider applications across a range of research areas. The paper describes the development and evaluation of four methods for semiautomated harvesting of geospatial language clauses from text to create a corpus of geospatial language. The most successful methods use a set of geospatial syntactic templates that describe common patterns of grammatical geospatial word categories, combined with extensible lists of members of those categories. The best method provides a maximum precision of 0.66, and is being used in an ongoing programme to harvest content from a range of web sites, followed by manual confirmation of content to create a monitor corpus. * Corresponding author. Email: [email protected] The stated aim of maximising the range of language collected (rather than focussing on the most common examples) across a range of English dialects presented challenges for the approach, as the wider range of included language also increased the rate at which text was incorrectly identified as geospatial when in fact it was not. The method was also limited in its ability to exclude metaphoric use of spatial language and sporting references. When run against actual web sites, the proportion of automatically harvested clauses that were manually confirmed as geospatial varied widely depending on the web source concerned, ranging from 5% to 78%, with an overall total of 42% for the first 31 sites harvested.