Leveraging Linked Data to Discover Semantic Relations within Data Sources. Mohsen Taheriyan Craig A. Knoblock Pedro Szekely Jose Luis Ambite

Transcription

1 Leveraging Linked Data to Discover Semantic Relations within Data Sources Mohsen Taheriyan Craig A. Knoblock Pedro Szekely Jose Luis Ambite

2 Domain Ontology CIDOC-CRM Source Map Structured Data to Ontologies Map the source to the classes & properties in an ontology title date name 1 The Island 2009 Walton Ford 2 Excavation at Night 1908 George Wesley Bellows 3 Rose Garden 1901 Maria Oakey Dewing 1

3 Semantic Types E35_Title E52_Time-Span E82_Actor_Appellation rdfs:label P82_at_some_time_within rdfs:label title date name 1 The Island 2009 Walton Ford 2 Excavation at Night 1908 George Wesley Bellows 3 Rose Garden 1901 Maria Oakey Dewing 2

4 Relationships E22_Man-Made_Object P108_was_produced_by P102_has_title E12_Production P14_carried_out_by E21_Person E35_Title E52_Time-Span E82_Actor_Appellation rdfs:label P82_at_some_time_within rdfs:label title date name 1 The Island 2009 Walton Ford 2 Excavation at Night 1908 George Wesley Bellows 3 Rose Garden 1901 Maria Oakey Dewing 3

5 Problem: How to automatically infer semantic relations?

6 Idea Exploit the relationships within already published linked data 5

7 Approach Input Output Target source (S) A ranked set of semantic Domain Ontologies (O) models for S Semantic labels of S Linked Data (in the same domain) Extract schema-level graph patterns from LD Construct a graph from LD patterns and the ontology Generate and rank semantic models 6

8 Approach Input Output Target source (S) A ranked set of semantic Domain Ontologies (O) models for S Semantic labels of S Linked Data (in the same domain) Extract schema-level graph patterns from LD Construct a graph from LD patterns and the ontology Generate and rank semantic models 7

9 Schema-Level LD Patterns../personinstitution/57551 rdf:type skos:preflabel E21_Person Thomas Burgon LD fragment from the British Museum P98i_was_born../personinstitution/57551/birth rdf:type E67_Birth../personinstitution/57551/birth/ date rdf:type rdfs:label E52_Time-Span

10 Schema-Level LD Patterns../personinstitution/57551 rdf:type skos:preflabel E21_Person Thomas Burgon LD fragment from the British Museum P98i_was_born../personinstitution/57551/birth rdf:type E67_Birth Pattern../personinstitution/57551/birth/ date rdf:type rdfs:label E52_Time-Span 1787 P98i_was_born E21_Person E67_Birth E52_Time-Span 9

11 Pattern Templates Many possible templates for patterns Example: patterns for classes C1, C2, C3 Consider only tree patterns Limit the length of the patterns 10

12 length 1 Extracting LD Patterns Use SPARQL to extract patterns of length one Person organizer Event Event location Place Person born Place 11

13 length 2 Extracting LD Patterns Iteratively construct larger patterns by joining with patterns of length 1 organizer location Person Event Place Event location Place born Person organizer Person born born Person born Event Place Place Place location Event location organizer Person organizer Place Place Event Event 12

14 Extracting LD Patterns Filter out the patterns not appearing in the data organizer location Person Event Place Event location Place born Person organizer Person born born Person born Event Place Place Place location Event location organizer Person organizer Place Place Event Event 13

15 Approach Input Output Target source (S) A ranked set of semantic Domain Ontologies (O) models for S Semantic labels of S Linked Data (in the same domain) Extract schema-level graph patterns from LD Construct a graph from LD patterns and the ontology Generate and rank semantic models 14

16 Merge the Patterns into a Graph Start from longer patterns, skip the ones already in the graph E22_Man-Made_Object P102_has_title P108i_was_produced_by P14_carried_out_by E53_Title E12_Production E21_Person E39_Actor P98i_was_born E52_Time-Span E67_Birth E82_Actor_Appellation 15

17 Weighting the Links Less weight for more popular links W = (1 - freq)/(total count of links) 0.70 P102_has_title E22_Man-Made_Object 0.87 P108i_was_produced_by 0.84 P14_carried_out_by E53_Title E12_Production E21_Person E39_Actor P98i_was_born E52_Time-Span E67_Birth E82_Actor_Appellation 16

18 Coherence Links from the same pattern have the same tag 0.70 P102_has_title m1 E22_Man-Made_Object 0.87 P108i_was_produced_by m P14_carried_out_by m5 E53_Title E12_Production E21_Person E39_Actor 0.95 m P98i_was_born m m m E52_Time-Span E67_Birth m3 E82_Actor_Appellation 17

19 Add the paths from the Ontology High weights for links that do not have any instance in the data 0.70 P102_has_title m1 E22_Man-Made_Object 0.87 P108i_was_produced_by m2 100 P14_carried_out_by 0.84 P14_carried_out_by m5 E53_Title E12_Production E21_Person E39_Actor 0.95 m P98i_was_born m m m E52_Time-Span E67_Birth m3 E82_Actor_Appellation 18

20 Approach Input Output Target source (S) A ranked set of semantic Domain Ontologies (O) models for S Semantic labels of S Linked Data (in the same domain) Extract schema-level graph patterns from LD Construct a graph from LD patterns and the ontology Generate and rank semantic models 19

21 Map Semantic Labels to the Graph 0.70 P102_has_title m1 E22_Man-Made_Object 0.87 P108i_was_produced_by m2 100 P14_carried_out_by 0.84 P14_carried_out_by m5 E53_Title E12_Production E21_Person E39_Actor 0.95 m P98i_was_born m m m E52_Time-Span E67_Birth m3 E82_Actor_Appellation 20

22 Map Semantic Labels to the Graph 0.70 P102_has_title m1 E22_Man-Made_Object 0.87 P108i_was_produced_by m2 100 P14_carried_out_by 0.84 P14_carried_out_by m5 E53_Title E12_Production E21_Person E39_Actor 0.95 m P98i_was_born m m m E52_Time-Span E67_Birth m3 E82_Actor_Appellation 21

23 Generate Semantic Models Compute top k minimal trees Consider both coherence and popularity 0.70 P102_has_title m1 E22_Man-Made_Object 0.87 P108i_was_produced_by m2 100 P14_carried_out_by 0.84 P14_carried_out_by m5 E53_Title E12_Production E21_Person E39_Actor 0.95 m P98i_was_born m m m E52_Time-Span E67_Birth m3 E82_Actor_Appellation 22

24 Evaluation Dataset Ontology Gold Standard Models 29 museum data sources 458 attributes (columns) 29 museum data sources 458 attributes 29 museum data sources 329 attributes 15 sources containing data about weapon ads 175 attributes CRM 147 classes 409 properties CRM 147 classes 409 properties EDM 147 classes 409 properties schema.org (ext) 736 classes 1081 properties 852 nodes 825 links 852 nodes 825 links 470 nodes 441 links 261 nodes 246 links Linked Data RDF generated from the same dataset (leave-one-out) RDF published by Smithsonian American Art Museum (more than 3 million triples) RDF generated from the same dataset (leave-one-out) RDF generated from the same dataset (leave-one-out) 23

25 Example Gold Standard Models 24

26 Evaluation Compute precision and recall (between learned links and correct links) Correct semantic labels are given location Artwork creator location Museum founder Museum Person Artwork Person correct model <Artwork,location,Museum> <Artwork,creator,Person> learned model <Museum,founder,Person> <Artwork,location,Museum> Precision: 0.5 Recall:

27 Results max len of patter ns Museum CRM (leave-oneout) Museum CRM (Smithsonian LD) Museum EDM Weapon schema.org precision recall precision recall precision recall precision recall Very low accuracy if only using the ontology paths Considering coherence improves the quality of the models (longer patterns increase the accuracy) Higher precision & recall for less complex ontologies 26

28 Related Work Understand semantics of Web tables [Wang et al., 2012] [Limaye et al., 2010] [Venetis et al., 2011] Link table values to the LOD entities [Muoz et al., 2013] [Mulwad et al., 2013] Learn semantic models from previously modeled sources (Karma) [Taheriyan et al, 2015] Extract schema-level patterns (SLPs, length one) from LOD [Schaible et al., 2016] E.g., ({Person,Player},{knows},{Person,Coach}) 27

29 Discussion Manually constructing semantic models is hard & expensive Needs domain knowledge and expertise in SW technologies Often requires many user interactions in modeling tools Infer semantic relations from linked data The suggested model can be refined in tools such as Karma Help to publish consistent RDF data 28