Detecting geospatial location descriptions in natural language text

References to geographic locations are common in text data sources including social media and web pages. They take different forms from simple place names to relative expressions that describe location through a spatial relationship to a reference object (e.g. the house beside the Waikato River). Often complex, multi-word phrases are employed (e.g. the road and railway cross at right angles; the road in line with the canal) where spatial relationships are communicated with various parts of speech including prepositions, verbs, adverbs and adjectives. We address the problem of automatically detecting relative geospatial location descriptions, which we define as those that include spatial relation terms referencing geographic objects, and distinguishing them from non-geographical descriptions of location (e.g. the book on the table). We experiment with several methods for automated classification of text expressions, using features for machine learning that include bag of words that detect distinctive words, word embeddings that encode meanings of words and manually identified language patterns that characterise geospatial expressions. Using three data sets created for this study, we find that ensemble and meta-classifier approaches, that variously combine predictions from several other classifiers with data features, provide the best F-measure of 0.90 for detecting geospatial expressions.