Strategies for gazetteer improvement and enrichment

Abstract

Many applications that use geographical databases (a.k.a. gazetteers) rely on the accuracy of the information in the database. However, poor data quality is an issue in gazetteers; often data is integrated from multiple sources with different quality constraints and there may not be much detail on the sources and the quality of the data. One major consequence of this is that the geographical scope of a location and/or its position may not be known or accurate. In this thesis, we develop novel strategies to accurately derive the geographical scope of places. Our strategies use the spatial hierarchy of a gazetteer as well as other public information (such as area) to construct a bounding box for each place. We present a probabilistic model of our approach and demonstrate the effectiveness of the bounding boxes in refining the spatial hierarchy of a gazetteer and augmenting it with other public data. Experimental evaluation on two public-domain gazetteers show that the proposed approaches significantly outperform, in terms of the accuracy of the bounding boxes, a baseline that is based on the parent-child relationship of a gazetteer. More specifically, our approaches outperform the baseline by 19-33% in terms of accuracy in a wide range of settings. Among applications, we show how these bounding boxes provide a generic way to improve the accuracy and usability of a gazetteer.