Automatic Semantic Transformation between Geo-Ontologies
Zusammenfassung der Projektergebnisse
The derivation of semantic transformation rules based on instances has many advantages: on the one hand, it opens the way to automation, on the other hand the instances represent already interpretations of the ontologies that are usually made from humans; thus the instances do reflect the diversity of possible interpretations of the ontologies. Furthermore, basing on the geometry of many instances, new geometric and topologic information can be derived to draw more inferences or improve the given semantic descriptions. E.g. instead of a specification of a minimal area the declaration of a detailed area interval is possible. We have presented a set of methods to explore the potential of geometric reasoning for extraction of semantic correspondences. The main results are: - Extraction of correspondence measures between object classes in two different datasets. The accuracy of the correspondence is measured statistically. - Development of matching approach for data in different scales and with possibly different geometric types. - Development of an approach to learn transformation rules between datasets of different granularity. In this way, relationships between two datasets could be determined. These rules can be used to generalize datasets for small scale applications. - Development of an approach for the classification of large spatial datasets based on geometric criteria and its application for the interpretation of a second dataset. Such an approach is very beneficial when it comes to the use of data which is collected by volunteers and which often lacks of clear semantics due to a missing object catalogue and/or different understanding of the meaning of objects. In our approaches we determined the quality of the correspondences using statistical measures. That means, we interpret the larger the common areas or the higher the number of matches are, the higher are also the class correspondences. Since our results are based on mere geometric analysis, we can only determine corresponding geometric features. Using our approach, it cannot be inferred, if the datasets have the same underlying semantics, e.g. one dataset can determine water quality of a river and the animals living in it, whereas the other describes the history and the river as traffic object. The only valid inference refers to the fact that both classes refer to the same physical objects, and that this physical object can be described with both class descriptions and ontologies. Using geometric criteria to assign the objects a certain semantic can help to give preliminary annotations to datasets which do lack a clear semantics, like Volunteered Geographic Information. These annotations could be termed a “light semantics”, as they are on the one hand based only on geometry and also there is a certain uncertainty in the classification. Still, however, it can be useful for different tasks. This is another issue, namely the question of the use of the results for certain applications. Depending on the given quality measure, different inferences can be drawn. E.g. if the algorithm determines a high quality of a correspondence, then this can be used to further applications like the transfer of attributes, as then, one can be sure to deal with the same object. If the correspondence is given only with a lower quality value, the subsequent tasks should take this lower quality into account.
