Improving the Performance of the DL-Learner SPARQL Component for Semantic Web Applications
- 1. Creating Knowledge out of Interlinked Data JIST 2012 – Page 1 http://lod2.eu Improving the Performance of the DL-Learner SPARQL Component for Semantic Web Applications Didier Cherix, Sebastian Hellmann, Jens Lehmann http://slideshare.net/kurzum http://dl-learner.org http://lod2.eu AKSW, Universität Leipzig LOD2 Presentation . 02.09.2010 . Page http://lod2.eu
- 2. JIST 2012 – Page 2 http://lod2.eu Motivation: 2007 - 2012 DL-Learner was developed in parallel to DBpedia at University Leipzig since 2007 DL-Learner is a tool for learning concepts in Description Logics (DLs) from user- provided examples. Worked very well for small to medium sized data sets, e.g. Carcinogenesis an other ML problems from the UCI ML repository Limit is the capacity of current OWL-DL reasoners Challenge was (and is) to do reasoning-based, supervized Machine Learning on the DBpedia Dataset (> 200 Mio triples) or larger datasets
- 3. JIST 2012 – Page 3 http://lod2.eu Introduction DL-Learner
- 4. JIST 2012 – Page 4 http://lod2.eu Introduction DL-Learner Very large search space Reasoner instance checks
- 5. JIST 2012 – Page 5 http://lod2.eu Introduction DL-Learner
- 6. JIST 2012 – Page 6 http://lod2.eu Introduction DL-Learner DL-Learner heavily relies on instance checks for machine learning, so the OWL Reasoner is the bottle neck Underlying idea: Only select relevant data for the Machine Learning Problem based on user-given examples → Reduces the amount of triples that have to be given to a reasoner → Reduces complexity and size of the OWL schema Brute-force approach: Load all data into the OWL Reasoner, then do instance checks → infeasible for Dbpedia Iterative approach (old component): Iterate over all instances and fetch the data recursively → inefficient even with caching
- 7. JIST 2012 – Page 7 http://lod2.eu Introduction DL-Learner
- 8. JIST 2012 – Page 8 http://lod2.eu Introduction DL-Learner
- 9. JIST 2012 – Page 9 http://lod2.eu Introduction DL-Learner
- 10. JIST 2012 – Page 10 http://lod2.eu Introduction DL-Learner
- 11. JIST 2012 – Page 11 http://lod2.eu Introduction DL-Learner
- 12. JIST 2012 – Page 12 http://lod2.eu Introduction DL-Learner Challenge: What is the most efficient way to retrieve such a fragment?
- 13. JIST 2012 – Page 13 http://lod2.eu Improvements of the New Component • Step 1: Indexing the T-Box: • Download the OWL Schema and index it in memory • either via SPARQL or OWL file
- 14. JIST 2012 – Page 14 http://lod2.eu Improvements of the New Component • Step 2: A-Box Queries Parameter recursion depth: Retrieve newly discovered bindings to ?o until a certain depth is reached.
- 15. JIST 2012 – Page 15 http://lod2.eu Improvements of the New Component • Step 3: Typing the retrieved instances
- 16. JIST 2012 – Page 16 http://lod2.eu Improvements of the New Component • Step 4: T-Box Index: All “relevant” T-Box information is added via the index to the fragment. For each class already in the fragment. all superclasses and their equivalentClass axioms are added
- 17. JIST 2012 – Page 17 http://lod2.eu Benchmarking - Speed For each class in DBpedia Ontology: - 30 instances as positives - 30 negatives from a sister class
- 18. JIST 2012 – Page 18 http://lod2.eu Benchmarking – F-Measure on the training data 70% of the results for each class had an F-measure of 90-100% on the training data
- 19. JIST 2012 – Page 19 http://lod2.eu SPARQL Retrieval Component Impact • DL-Learner – http://dl-learner.org • DBpedia Navigator • Tiger Corpus Navigator • AutoSPARQL - http://autosparql.dl-learner.org/ • HANNE – http://hanne.aksw.org • ORE - http://aksw.org/Projects/ORE Sebastian Hellmann, Jens Lehmann und Sören Auer: Learning of OWL Class Descriptions on Very Large Knowledge Bases In: International Journal on Semantic Web and Information Systems, 2009 Web Applications Active Learning → User Interaction and Feedback
- 20. JIST 2012 – Page 20 http://lod2.eu Future Work • Research Paper in Session 4b (tomorrow at 15:10) Navigation-induced Knowledge Engineering by Example • Caching + more sophisticated options • Large scale learning problems http://slideshare.net/kurzum Homepage: http://dl-learner.org Source code: http://sourceforge.net/projects/dl-learner/
- 21. JIST 2012 – Page 21 http://lod2.eu Example Sebastian Hellmann, Jens Lehmann, Jörg Unbehauen, Claus Stadler, Thanh Nghia Lam und Markus Strohmaier: Navigation-induced Knowledge Engineering by Example In: JIST 2012
- 22. JIST 2012 – Page 22 http://lod2.eu Example Sebastian Hellmann, Jens Lehmann und Sören Auer: Learning of OWL Class Descriptions on Very Large Knowledge Bases In: International Journal on Semantic Web and Information Systems, 2009