To Learn or to Rule: Two Approaches for Extracting Geographical Information from Unstructured Text

Geographical data plays an important role on the Web: recent search engine statistics regularly confirm that a growing number of search queries have a locale context or contain terms referring to locations. Assessing geographical relevance for Web pages and text documents requires information extraction techniques for recognizing and disambiguating geographical entities from unstructured text. We present a new corpus for evaluation purposes, which we make publicly available for research, describe two approaches for extracting geographical entities from English text—one based on heuristics, the other relying on machine learning techniques—and perform an extensive discussion of those two approaches. Furthermore, we compare our approach to other publicly available location extraction services. Our results show, that the presented approaches outperform current state of the art systems.