Ontology Alignment, Matching and TranslationIn the old daysOntology matchingThe Semantic Web VisionBut…Discovering MappingsUsing MappingsSlide 8Let’s do this as a project?OntologyAlignment,Matching and TranslationIn the old daysPeople have been building knowledge based systems for ~40 yearsThere was not much interest in integrating them before the mid 80sCyc argued (~1985) for the utility of having a shared KB, but just one that all would refer toAgent oriented approaches in the 90s imagined having multiple share ontologies–KIF was proposed as an interlingua for importing and exporting knowledgeOntology matchingMatching or aligning knowledge encoded in different KR languages can be very hardDifferences in the KR languages can be major or subtle and both can cause problems–E.g., FOL, vs. bayesian vs defaults vs sterotypes vs …Trying to deal with this problem usually means that you need to adopt a very abstract and flexible interlinguaIt’s much easier if we can limit ourselves to translation between different schemas in the same KR languages–e.g., like the problem of schema mapping in RDBMsThe Semantic Web VisionEveryone uses the same Knowledge Representation language – OWLThere is no assumption of having ONE ontology for any topic–Assume many will be used and invest in techniques for translation–Analogy for how the UN manages translationsOWL also has primitives that can describe some mappings–foaf:Person owl:sameClassAs wn:Human–wn:Human rdfs:subClass spire:homoSapienBut…Mappings can be complex–o1:Boy = intersection(o2:Human, o2:Male, complement(o2:Adult))–Here’s where DL can help and do so efficientlyNot all useful mappings can be expressed in FOLo1:Mammal ~ o2:FurryAnimal–Dolphins are mammals but are not furry–We would benefit from conditional probabilities, e.g., p(o1:Mammal|o2:FurryAnimal) and p(o2:FurryAnimal|o1:Mammal)Peng and others are exploring this ide–Probabilities can come from human judgments or shared data–Need to respect the FOL constraints inherent in OWLDiscovering MappingsAutomatically discovering the mappings at a schema level–Hard problem without common instance dataSemi-automatically discovering the mappings at a schema level–Can use OWL’s constraints, e.g., if a:C1<a:C2 and b:C3<b:C4, then b:C4<a:C1 implies b:C3<A:C1 and b:C3<a:C2Using instance data to suggest or rule out alignments–If we’re lucky, the ontologies might share some instances–We might also note patterns (e.g., “138-35-9866”) in literal dataWe can also get the mappings manually or collect them using SwoogleUsing MappingsOnce we have the mappings, how do we use them?One model for translation: merge the ontology and instance data from the source data and the ontology from the target ontologyAdd bridging axioms for source and target ontologies–o1:Boy = intersection(o2:Human, o2:Male, complement(o2:Adult))–o3:Journal < o4:SerialDraw all possible interferences over the instance dataWrite out the instance data expressed in the target ontologiesUsing MappingsSuch systems have been built–Dejing Dou, Drew McDermott, and Peishen Qi “Ontology translation by ontology merging and automated reasoning”. In Proc. EKAW Workshop on Ontologies for Multi-Agent Systems. 2002.–http://cs-www.cs.yale.edu/homes/dvm/papers/DouMcDermottQi02.pdfAnd the approach may be used in many ad hoc, one-off translation systemsBut no widely used tools are available, to my knowledgeLet’s do this as a
View Full Document