Semantic Similarity based on Weighted Ontology - Geographical Information Systems: Trends and Technologies - page 288

al. 2011; Sánchez et al. 2010; 2012). An Ontology is an explicit and formal
specifi cation of sharing information (Gruber 1993). Generally, ontologies
in GISs constitute the hierarchy structures, which can be derived from
appropriate classifi cation of data in geographical information domain.
Some proposals in the GIS domain use a given reference ontology in the
semantic similarity measuring process, for instance in (Rodriguez and
Egenhofer 2003; 2004; Formica and Pourabbas 2009). In particular, in
(Formica and Pourabbas 2009) the authors propose a semantic similarity
method, called GSim , which is conceived by combining the similarity of
geographical concepts organized as a Part-of taxonomy and the feature (or
attribute) similarity of geographical classes. To this end, the information
content approach of Lin (1998) has been extended and integrated with a
method for attribute similarity, which is based on the maximum weighted
matching problem in bipartite graphs (Kuhn 1955). The information content
of a concept c is computed according to the following expression: - logw ( c ),
where w ( c ) is the weight of the concept. The GSim method, as large majority
of proposals in the literature, indicates the use of weights of concepts (or
geographical classes) derived from WordNet (Miller 1995) frequencies.
WordNet (a lexical ontology for the English language) provides, for a given
concept (noun), the natural language defi nition, hypernyms, hyponyms,
synonyms, etc., and also a measure of the frequency of the concept, by using
noun frequencies from the Brown Corpus of American English (Francis
and Kucera 1979). However, weights are often not available for all possible
concepts (as discussed in the section devoted to weight assignment).
In this chapter, we focus on the problem of weight assignment to
concepts. We propose and formalize a probability - based approach as an
alternative to the frequency-based measures provided by WordNet. This
approach is characterized by three notions, called uniform probabilistic
approach , uniform probabilistic weighted arc and non - uniform probabilistic
weighted arc . They essentially differ on the basis of nodes, arcs and depth
of the reference ontology. Successively, we describe in detail the selected
methods, Lin (1998), Dice (Maarek et al. 1991), Matching-Distance Similarity
Measure—for short MDSM (Rodriguez and Egenhofer 2003; 2004), and the
GSim method as representative similarity methods in the literature. As
anticipated, Lin method is an information content-based approach, Dice and
MDSM methods are based on the feature-based approach, and the GSim
method is a combined approach, which is conceived to capture both the
concept similarity within the hierarchy, and the attribute similarity (or tuple
similarity) of geographic concepts or classes. In the section of experimental
analysis, these methods are investigated in detail and the results of the
contrast of the selected methods are illustrated. An experiment about two
different weight assignment approaches is provided and their impact on
the selected methods is illustrated.