Semantic Intensity Spectrum and Semantic Integration Algorithms
- 1. Semantic Intensity Spectrum At least 25 systems have been developed for ontology alignment, matching. A classification technique for ontology alignment approaches Based on semantic intensity Semantics: the intended meanings of ontological entities Some methods consider only syntactical features (semantic poor) E.g. String distance, String equality Semantics are added via: meanings of words provided by external lexicons Positions in taxonomies Relations with other ontological entities Logic entailment Classified instance data Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 2. Semantic Intensity Spectrum SIS Diagram http://www.aktors.org/crosi/si-spectrum/ Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 3. Semantic Intensity Spectrum Existing systems Duplicate efforts in developing largely overlapped algorithms Re-implement algorithms e.g. string distance Similar heuristic rules Different performance and different results w.r.t. the same test sets Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 4. Semantic Intensity Spectrum Diversity of existing systems Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 5. Semantic Intensity Spectrum A possible solution Combine existing systems to minimise development efforts Possibility of combining in a meaningful way Many systems output in compatible format Heterogeneous outputs need to be normalised using heuristic rules, e.g. convert “more general than” into numeric values Reuse available packages E.g. SecondString for computing string distance Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 6. A principled architecture Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 7. A principled architecture: Signature Extraction Ontologies can be captured with a set of ontological signatures Local signatures: Labels, IDs, and URIs Declaimed properties, property domains and ranges Equivalent and complement classes, inverse and functional properties Instantiated classes Global signatures: Super-, sub-classes, properties Disjoint classes Sibling classes Comments, version information Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 8. A principled architecture: Multiple matchers Specialised internal matchers targeting at particular signatures Name matchers String distance based matchers WordNet based matchers Class matchers Taxonomy based matchers Definition based matchers Invoking existing ontology matching/alignment systems as external matchers FOAM API INRIA Alignment API Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 9. CMS design commitments Avoid reinventing the wheel Use existing packages to enhance internal matchers Use existing mapping/alignment systems as external matchers Semantically enriched matchers based on the definition of concepts Propagate similarity along concept hierarchies Refine concept similarity by taking into account the names, domains and ranges of declared properties Compute similarity using WordNet hierarchies Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 10. String distance Reuse of existing packages SecondString Metrics Jaro, MongeElkan, NeedlemanWunsh, etc. Soundex Metrics Consider only the local names of ontological entities Namespace is ignored Names of super(sub)classes are ignored Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 11. WordNet-based algorithms Use JWNL WordNet Java Lib Names only Synonyms are retrieved and compared with string equality or string distance Composite names are split and stop words are removed E.g. “has_name” => “name” WordNet hierarchy Calculate distance between two Words Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 12. WordNet-based algorithms WordNet hierarchy h the distance between “Word” and Root h’ the distance between “Word’” and Root H the distance between common subsumer of “Word” and “Word’” and Root Similarity between “Word” and “Word’” is computed as 2H/(h+h’) Root Common Subsumer Word’ Word Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work H h’ h
- 13. Canonical Name C=> A.B.D.C C’=> A’.B’.D’.E’.C’ Compute the similarity between C and C’ as well as the respective similarity between every pair of super classes of C and C’ Penalise the similarity between C and C’ with those of their super classes C’ A’ B’ D’ E’ A B C D Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 14. Structure algorithm f (name similarity, domain similarity, range similarity ) Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work P1’(C’, B’) P2’ P3’ P1(C, B) P2 P3 H G Domain of P1 H’ G’ I’ Domain of P1’ A B D Range of P1 E A’ B’ D’ E’ Range of P1’ F’
- 15. Structure algorithm cnt’d Structure Retrieve declaimed properties For each property, retrieve its domains and ranges Compare property’s name, domain and range StructurePlus Compare also the super and sub classes of property’s domain and range When compare domains and ranges, using existing name matching techniques Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 16. Implemented Matchers powered by existing java libs Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 17. Post alignment Aggregator Weighted average based aggregation Weights are manually set by users Evaluator Nothing is more qualified than a human inspector with domain knowledge and experiences Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 18. CMS: deployment options Run CMS from command line A batch file is provided Invoke CMS as an API Run CMS as a service via JSP interface Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 19. Demo Ontologies: web directory, small size and simple structure Run with different weights Output to different formats OWL, SKOS, HTML, XML(OAEI) Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work
- 20. Use of SIS Declaimed functionalities of existing systems can be justified against this spectrum A reference for selecting the right mapping techniques for a particular problem A designer’s aid for navigating through different mapping approaches with emphasis on the use of semantics Project aims & targets Timeline and deliverables Semantic Intensity Spectrum Modular architecture Algorithms CMS Evaluation Lessons learnt & future work