Access to Earth Observation (EO) products remains di cult for end-users. To address this, we developed the Prod-Trees platform1[2], a semantically enabled search engine for EO products. Users guide their search through a number of ontologies related to EO domain. To facilitate users in nding terms that t better to their needs, we created mappings between these ontologies. In this paper, we present Pythia, an ontology matching system that utilizes and combines various matching techniques [1,3,4] to create mappings between two ontologies. Pythia is a combination of a string-based technique utilizing Apache Lucene's features, a language-based technique based on WordNet, and a graph-based technique that uses the structure of the ontology and the mappings produced by the two previous techniques. The system supports SKOS ontologies. Therefore, the mappings are also expressed in SKOS using the de ned properties for matching concepts: skos:exactMatch, skos:relatedMatch, skos:broadMatch, and skos:narrowMatch. Based on these, we create four di erent types of mappings. A terminological matcher is responsible for implementing the string- and language-based techniques, both applied on the concepts labels (skos:prefLabel, skos:altLabel and skos:hiddenLabel ). The mappings created by this component can either be skos:exactMatch or skos:relatedMatch. The string-based technique uses Lucene for indexing and searching. With Lucene, one can create documents and add elds of a speci c type to these documents. When searching the documents, the user can specify which eld he wants to search. Taking advantage of Lucene capabilities, the terminological matcher indexes the target ontology. A new document is created for each concept and each available property of the concept is added as a new eld. String normalization functions are applied to the eld and unnecessary stop words are removed. When searching for concepts similar to concept A (from the source ontology), the prefLabel, altLabel, and hiddenLabel elds of the indexed ontology are searched using the prefLabel of concept A. The search results fetched back, are ranked according to the string similarity of the compared strings (e.g., skos:prefLabel of A and the prefLabel eld of a document). This is feasible due to the string similarity functions implemented in Lucene. Also, since each eld is indexed, only the index of the speci ed eld is searched, and not all the concepts. Lucene returns multiple related results. If the two strings are the same, a skos:exactMatch is created between A and the corresponding concept from the ? This work was supported by the Prod-Trees project funded by ESA ESRIN. 1 A video demonstrating the functionalities of the Prod-Trees platform is available at http://bit.ly/ProdTreesPlatform.
target ontology. Otherwise, and only if one string is a substring of the other (e.g., \Elevation" and \Digital Elevation Model"), a skos:relatedMatch is created.
The language-based technique uses WordNet, a lexical database for English. The technique is optional and can be bypassed, as it adds noise to the results. Putting WordNet to use, a new eld, called relLabel, is created in the Lucene document of each concept. relLabel enhances each concept's labels, by adding synonyms and other related words found in WordNet. During the search, the relLabel elds of the documents are searched, and if a similarity is discovered, a skos:relatedMatch relation is created between the corresponding concepts.
In case there are concepts from the source ontology with no skos:exactMatch mappings, a structural matcher is invoked. This component implements a graph-based technique creating either skos:narrowMatch or skos:broadMatch mappings. Taking as input a concept A from the source ontology, the matcher nds all the broaders and narrowers of A. Afterwards, it checks whether a skos:exactMatch was created by the terminological matcher for one of these concepts. If it did, then a new mapping can be derived. For example, if a skos:exactMatch exists between concept B (which is a broader of A) and concept B'(from the target ontology), then it can be derived that B' will be a skos:broadMatch of A. Similarly, we can create a skos:narrowMatch.
The matcher also checks whether the concepts B and N hold skos:narrowMatch or skos:broadMatch relations with concepts from the target ontology. If a skos:broadMatch exists between B and a concept B", then it is safe to conclude that B" will also be a skos:broadMatch of A. This means that when a skos:broadMatch exists between a concept B from the source ontology and a concept B" from the target ontology, then this relation can be propagated to concept's B narrowers. Similarly, a skos:narrowMatch between a concept N and a concept N", can be propagated to concept's N broaders. In any other case, no mappings can be derived. When all the concepts are examined, if new mappings were created by the structural matcher, the described process is repeated. Otherwise, Pythia proceeds with the exportation of the mappings to RDF.
Despite the simplicity of the techniques, the results are quite satisfying. Especially, the performance of the language-based technique, which allows tuning WordNet. By stating the types of relations WordNet discovers for a given word, it gives control over the percentage of valid mappings. A higher degree of trust for the nal results can be gained with extensions such as a user-evalutation process and the use of domain-speci c vocabularies coupled with Wordnet.