A Hybrid Semantic Similarity Measure for Spatial Information Retrieval

Semantic similarity is central to many cognitive processes and plays an important role in the way humans process and reason about information. In particular, the retrieval of knowledge from memory hinges crucially on similarity. Likewise, information retrieval systems use similarity to detect relevant information for a given query. Current information retrieval systems apply mainly syntactic techniques to determine similarity. Although such syntactic similarity measures have performed strongly with resources containing large amounts of text, they cannot appropriately cope with syntactic and semantic heterogeneity and ambiguity, if the semantics of the terms is not explicitly available. Therefore, they are rather rigid and inflexible as they cannot adapt to the user's requirements and conceptualization of the domain. Furthermore, geographic features are distinguished via their geometric and thematic data. It is often not possible to capture the complex semantics of geographic features by a single name or a textual description. Therefore spatial data is different from text documents typically found in enterprise databases or on the Web. Retrieval of spatial information requires new, intelligent retrieval mechanisms that satisfy its specific requirements. A semantics-based solution can more easily adapt to user needs and therefore increases the flexibility and usability of spatial data and retrieval methods. This paper investigates the suitability of various approaches-originally developed to explain human similarity judgement-in the context of spatial information retrieval. We propose a new, hybrid approach for semantic similarity measurement, which can represent the complex semantics of spatial data. It allows for retrieving relevant data by determining the similarity between the query and the semantic descriptions of geographic feature types within the database. The hybrid similarity measure combines the geometric structure of conceptual spaces with the relational structure of semantic nets to one, cognitively plausible knowledge representation with an inherent similarity measure.