| Semantic similarity plays an important role in information systems by allowing the identification and management of objects that are conceptually close. As such, it constitutes a basis for designing mechanisms for information extraction and integration that satisfy users' needs. In geographic information systems, similarity assessment has so far been focused on geometric characteristics, which are treated in isolation from semantics. Although computer scientists have been working on models for semantic similarity assessment, these models are limited with respect to handling the cognitive properties of similarity, such as asymmetry and context dependence.; This thesis defines a computational model for semantic similarity among spatial entity classes, called the Matching-Distance model. Entity classes are modeled in an ontology by their distinguishing features (i.e., parts, functions, and attributes) and their semantic interrelations (i.e., is-a and part-whole relations). The Matching-Distance model combines the advantages of a matching process with a semantic-distance determination such that asymmetric values of similarity are obtained for entity classes that have different degrees of generalization. The model introduces contextual information to determine the relevance of distinguishing features in the matching process. This contextual information is specified by the user's intended operations and is used to define the set of entity classes in the domain of an application. Based on an extension of the Matching-Distance model, the Triple Matching-Distance model evaluates semantic similarity between entity classes across independent ontologies. This model compares entity classes in terms of the common components of the ontologies' specifications (i.e., lexicons, features, and semantic relations).; The main result of the thesis is that the Matching-Distance model corresponds to people's judgments of semantic similarity. The model performs well for evaluations among entity classes that are distinguished by their parts, functions, or attributes. By integrating contextual information, the Matching-Distance model improves its performance; however, the major determinant of the correctness of the model is a complete and detailed definition of entity classes. While lexicons and semantic relations are shown to be good indicators for identifying equivalent entity classes across ontologies, features are important components for detecting similarity among entity classes. |
