The work presented in this paper extends our work on matching formal and
informal geospatial ontologies [
Di erent geospatial instances may have the same purely lexical information and the same classi cation in terms of an ontology, but di erent locations. For example, there may be several restaurants called `Prezzo Ristorante' in the same city. Therefore, when matching geospatial instances, it is essential to use location information. However, few of existing ontology matching or data interlinking methods can match spatial instances e ectively.
There is also a choice in representing objects such as shopping centres as a collection of parts or as a single instance. For example, Victoria Centre is represented as a collection of shops and other businesses in OSGB and as a single instance in OSM. In order to produce a meaningful correspondence between instances in OSGB and in OSM, we propose to use `partOf' relation (mereological partOf in geometry and having similar labels). If for two instances a and b we get `partOf' relations in both directions, we generate a hypothesis that a and b belong to `sameAs'.
Here we propose a new method for establishing `sameAs' and `partOf' relations between geospatial instances from di erent ontologies. In crowd-sourced data, there is an increased possibility of error in measurement, and tendency to simplify shapes of buildings. For this reason, in our method we use bu ers to compare geometries of instances. The size of the bu er (intuitively, `the margin of error' ) is a parameter which can be arrived at experimentally or within reason set arbitrarily. Intuitively the optimal value of corresponds to the maximal deviation between two representations of the same object in two di erent data sets. In the case of OSGB and OSM data for Nottingham, this is experimentally determined as 20m. The new method has four steps.
Step 1: Extracting geometry sets. An ABox contains facts (geometry, lexical and semantic classi cation information) about geospatial instances. We extract a set of geometries, Gi, from all the spatial instances in each ABox Ai, i = 1; 2.
Step 2: Matching geometry sets. For two sets of geometries, G1; G2, a level of tolerance and tolerance for the second best , we generate the best two candidate matches for each geometry in G1 if they exist in G2, and the best two candidate matches for each geometry in G2 if they exist in G1. The candidates are selected by comparing minimal bu ers. The bu er of a geometry g, bu er (g; ) = fp : 9p0 2 g: distance(p; p0) g; ( > 0). For two geometries g and h, the minimal bu er of h containing g is bu er (h; ) such that g bu er (h; ) and for all 0 < , g 6 bu er (h; 0). For any geometry g, the minimal bu er ( ) of its best candidate o1 (g bu er (o1; )) is the smallest among those of all the candidates. We generate a 'bu ered part of' (BP T ) relation between each geometry and its candidates, i.e. (g; o1) 2 BP T ( ).
Step 3: Comparing labels. We use string comparison, including equality, inclusion, abbreviation and edit distance to check whether the labels, such as names or addresses, of two instances are similar. If there is no pair of labels of spatial instances s1; s2 that are similar, then their lexical information is incompatible, (s1; s2) 2 LF . Otherwise, their lexical information is compatible, (s1; s2) 2 LT .
For every pair of spatial instances s1; s2, if (g1; g2) 2 BP T ( ) (where gi is the geometry of si, i = 1; 2) and (s1; s2) 2 LT , then (s1; s2) 2 partOf possibly holds, and we will add it to the initial instance mapping M .
Step 4: Verifying initial instance mapping M using semantic classi cation
information. It is part of ontology (ABox and TBox) matching process, presented
in [
We implement the method described above as part of GeoMap [