Resolving fine granularity toponyms: Evaluation of a disambiguation approach

Landscape descriptions in natural language, for instance from historic corpora, are a complementary source to empirical ethnographic work, for example to research exploring variation in the use of basic levels or basic terms within landscapes across localities (c.f. Mark and Turk 2003, Burenhult and Levinson 2008, Turk et al. 2011), on the condition that such descriptions can be linked to space. A key challenge in linking language to space is the detection and resolution of toponyms (Purves and Jones 2008). Central to toponym resolution is the identification of a single unambiguous referent for a given toponym, which requires that toponym referent ambiguity is resolved (c,f, Amitay et al. 2004), i.e. does the document refer to London, England or London, Ontario. Some common state of the art approaches to toponym disambiguation use default rules, such that the most prominent referent location is resolved (c.f. Purves et al. 2007), population counts, also reflecting the prominence of referent locations (c.f. Martins et al. 2010) and geometric minimality, assuming that the areal footprint of a document is to be minimised (c.f. Leidner 2004). Leidner (2007) argued that toponym disambiguation had focused on populated places, since such locations are important for a variety of applications (e.g. local search or news mapping). However, if we wish to resolve toponyms with a fine spatial granularity, such as those typically used to reference mountains, hills, fields or hamlets in natural landscape descriptions, state of the art disambiguation approaches must be adapted to work independently from a priori toponym knowledge that is usually attached to populated places and commonly found in gazetteers (Hill 2006). We present an approach for toponym disambiguation, working independently from a priori toponym knowledge. We evaluate its performance over a baseline disambiguation technique on an extensive corpus consisting of 150 years of Swiss alpine literature (Volk et al. 2009). Toponym knowledge is gathered from geomorphometric characteristics at locations of toponyms. This reflects the strong relation between toponyms and topography, since toponyms are used to name geographic objects that are attached to the earth’s surface (Smith and Mark 2003) and therefore can be hypothesised to be bound to its characteristics. We show that in a user-centered evaluation with spatial queries (i.e. classifying articles contained in the alpine corpus as being relevant or not for certain spatial extents) our approach to disambiguation, using geomorphometric information, significantly outperforms baseline disambiguation (27% improvement).