Serving DBpedia with DOLCE - More Than Just Adding a Cherry on Top
Download as PPT, PDF0 likes769 views
The document discusses the integration of the DOLCE ontology with DBpedia to enhance its consistency checking and enrich its ontology. It highlights systematic errors found in DBpedia through clustering techniques, identifying prevalent issues such as overcommitment and misalignment of classes. The findings suggest that DOLCE significantly aids in pinpointing inconsistencies and could lead to improvements in future DBpedia versions.
Serving DBpedia with DOLCE - More Than Just Adding a Cherry on Top
- 1. Serving DBpedia with DOLCE More than Just Adding a Cherry on Top DBpedia Extraction Framework Mappings Wiki DOLCE Heiko Paulheim and Aldo Gangemi
- 2. 10/13/15 Heiko Paulheim and Aldo Gangemi 2 DOLCE Mappings: Served Since DBpedia 2014
- 3. 10/13/15 Heiko Paulheim and Aldo Gangemi 3 More than Just a Cherry on Top • DOLCE adds a layer of formalization – high level axioms – additional domain and range restrictions – fundamental disjointness (e.g., physical object vs. social object) • Enriches DBpedia ontology • Can be used for consistency checking
- 4. 10/13/15 Heiko Paulheim and Aldo Gangemi 4 More than Just a Cherry on Top Tim Berners-Lee Royal Society Award award range Description subclass of Social Agent disjoint with DBpedia ontology DBpedia instances DOLCE ontology Organisation is a Social Person equivalent class subclass of
- 5. 10/13/15 Heiko Paulheim and Aldo Gangemi 5 DBpedia in a Nutshell • Raw extraction from Wikipedia infoboxes • Infobox types and keys mapped to ontology – Crowdsourcing process (aka Mappings Wiki) – 735 classes – ~2,800 properties • Almost no disjointness – only 24 disjointness axioms – many of those are corner cases • MovingWalkway vs. Person f
- 6. 10/13/15 Heiko Paulheim and Aldo Gangemi 6 DOLCE in a Nutshell • A top level ontology • Defines top level classes and relations • Including rich axiomatization
- 7. 10/13/15 Heiko Paulheim and Aldo Gangemi 7 DOLCE in a Nutshell • Original DOLCE ontologies were too heavy weight – thus: hardly used on the semantic web – remember: a little semantics goes a long way! • DOLCE-Zero – Simplified version – Contains both DOLCE and D&S (Descriptions and Situations) – D&S introduces some high level design patterns
- 8. 10/13/15 Heiko Paulheim and Aldo Gangemi 8 Systematic vs. Individual Errors in DBpedia • Systematic errors – occur frequently following a pattern – e.g., organizations are frequently used as objects of the relation award • are likely to have a common root cause – wrong mapping from infobox to ontology – error in the extraction code – ... Tim Berners-Lee Royal Society Award award range Description subclass of Social Agent disjoint with DBpedia ontology DBpedia instances DOLCE ontology Organisation is a Social Person equivalent class subclass of
- 9. 10/13/15 Heiko Paulheim and Aldo Gangemi 9 Overall Workflow • Overall workflow • For each statement – add the statement plus all subject/object types to the ontology – check consistency – present inconsistent statements and explanations for inspection ; Reasoning 2 RDF Graph Statements + Types Inconsistent Statements + explanations User
- 10. 10/13/15 Heiko Paulheim and Aldo Gangemi 10 Overall Workflow • Inspecting a single statement takes 2.6 seconds – on a standard laptop • DBpedia 2014 has 15,001,543 statements – in the dbpedia-owl namespace → consistency checking would take 451 days • Solution – cache results for signatures (predicate + subject types + object types) – there are only 34,554 different signatures! ; Reasoning 2 RDF Graph Statements + Types Inconsistent Statements + explanations UserCache
- 11. 10/13/15 Heiko Paulheim and Aldo Gangemi 11 Overall Workflow • Overall, we find 3,654,255 inconsistent statements (24.4%) – cf.: only 97,749 (0.7%) without DOLCE • Too much to inspect! – We are looking for systematic errors – Cluster explanations w/ DBSCAN – Each cluster represents a systematic error ; Reasoning Clustering 2 RDF Graph Statements + Types Inconsistent Statements + explanations Clusters UserCache
- 12. 10/13/15 Heiko Paulheim and Aldo Gangemi 12 Clustering Inconsistent Statements • Each explanation as a binary vector – with 0/1 for all axioms involved in the explanations – 1,467 axioms (dimensions) in total • DBSCAN – Manhattan distance (i.e., number of axioms added/removed) – MinPts=100 (minimum frequency for a systematic error) – ε=4 (explanations in a cluster differ by two axioms at most)
- 13. 10/13/15 Heiko Paulheim and Aldo Gangemi 13 Major Systematic Errors • Inspection of the top 40 clusters – they contain 96% of all inconsistent statements • Overcommitment (19) – using properties in different contexts – e.g., dbo:team is defined as a relation between persons and sports teams – but also used for relating participating teams to events – fix: relax domain/range constraints, or introduce new properties • Metonymy (11) – ambiguous language terms – e.g., instances of dbo:Species contain both species as well as single animals – hard to refactor
- 14. 10/13/15 Heiko Paulheim and Aldo Gangemi 14 Major Systematic Errors • Misalignment (5) – classes/properties mapped to the wrong concept in DOLCE – occasionally occurs if intended and actual use differ – e.g., dbo:commander is more frequently used with events (e.g., battles) than military units – d0:hasParticipant rather than d0:coparticipatesWith – fix: change alignment • Version branching (3) – semantics of dbo concepts have changed – e.g., dbo:team in DBpedia 3.9: career stations and teams, in DBpedia 2014: athletes and teams – fix: change alignment
- 15. 10/13/15 Heiko Paulheim and Aldo Gangemi 15 ` A Look at the Long Tail • DBSCAN identifies clusters and “noise” – i.e., statements that are not contained in clusters • Manual inspection of a sample of 100 instances – 64 are erroneous – 30 are false negatives (i.e., correct statements) – 6 are questionable • Typical error sources in the long tail – are expected to be cross-cutting, i.e., occurring with various classes and properties
- 16. 10/13/15 Heiko Paulheim and Aldo Gangemi 16 A Look at the Long Tail • Typical error sources in the long tail • Link in longer text (23) – e.g., dbr:Cosmo_Kramer dbo:occupation dbr:Bagel . • Wrong link in Wikipedia (9) – e.g., dbr#Stone_(band) dbo:associatedMusicArtist dbr#Dementia . – Dementia should link to the band, not the disease • Redirects (7) – e.g., dbpedia#Ben_Casey dbo:company dbpedia#Bing_Crosby . – The target Bing_Crosby_Productions redirects to Bing_Crosby • Links with Anchors (6) – e.g., dbr:Tim_Berners-Lee dbo:award dbr:Royal_Society . – the original link target is Royal_Society#Fellows – anchors are ignored by DBpedia
- 17. 10/13/15 Heiko Paulheim and Aldo Gangemi 17 Conclusions • We have shown that – DOLCE helps identifying inconsistent statements – Cluster analysis allows for identifying systematic errors – User interaction is minimized • we analyzed one statement each from 40 clusters • corresponding to 3,497,068 affected statements • Outcomes for future DBpedia versions – DBpedia ontology changes – Mapping changes – DOLCE alignment changes – Bug reports
- 18. Serving DBpedia with DOLCE More than Just Adding a Cherry on Top DBpedia Extraction Framework Mappings Wiki DOLCE Heiko Paulheim and Aldo Gangemi