A Hitchhiker s Guide to Ontology

Transcription

1 A Hitchhiker s Guide to Ontology Fabian M. Suchanek Télécom ParisTech University, Paris, France

2 2 Fabian M. Suchanek money freedom permanent

3 3 Fabian M. Suchanek 2003: BSc in Cognitive Science University of Osnabrück/DE 2005: MSc in Computer Science Saarland University/DE 2008: PhD in Computer Science Max Planck Institute/DE money freedom permanent

4 4 Fabian M. Suchanek 2003: BSc in Cognitive Science University of Osnabrück/DE 2005: MSc in Computer Science Saarland University/DE 2008: PhD in Computer Science Max Planck Institute/DE money freedom permanent

5 5 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US money freedom permanent

6 6 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US money freedom permanent

7 7 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US 2010: PostDoc INRIA Saclay/FR money freedom permanent

8 8 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US 2010: PostDoc INRIA Saclay/FR money freedom permanent

9 9 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US 2010: PostDoc INRIA Saclay/FR 2012: Research group leader Max Planck Institute/DE money freedom permanent

10 10 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US 2010: PostDoc INRIA Saclay/FR 2012: Research group leader Max Planck Institute/DE money freedom permanent

11 11 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US 2010: PostDoc INRIA Saclay/FR 2012: Research group leader Max Planck Institute/DE 2013: Maître de Conférences Télécom ParisTech/FR money freedom permanent

12 12 Fabian M. Suchanek 2009: PostDoc at Microsoft Research Silicon Valley/US 2010: PostDoc INRIA Saclay/FR 2012: Research group leader Max Planck Institute/DE 2013: Maître de Conférences Télécom ParisTech/FR money freedom permanent

13 13 Fabian M. Suchanek I am an Elvis fan! Elvis, when I need you, I can hear you! Will there ever be someone like him?

14 14 Searching with Google Another singer called Elvis, young

15 15 Searching with Google Another singer called Elvis, young Carol Connors talks about her first boyfriend Carol Connors talks about her first boyfriend - a young singer called Elvis Presley.

16 16 Searching with Google Another singer called Elvis, young Younger Elvis, singer

17 17 Rephrasing does not help Another singer called Elvis, young Younger Elvis, singer Google, you don t understand! I want

18 18 We need structured knowledge Singer type type born 1935? born? To answer the question, the computer would need structured knowledge: an ontology

19 19 I work on Ontologies Singer type type born? born?

20 20 I work on Ontologies Constructing ontologies Singer type type born? born?

21 I work on Ontologies Constructing ontologies Singer type type born? born? Mining ontologies A B C 21

22 I work on Ontologies Constructing ontologies Singer type type born? born? Mining ontologies A B C Aligning ontologies 22

23 I work on Ontologies Constructing ontologies Singer Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 23

24 I work on Ontologies Constructing ontologies Singer Applying ontologies Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 24

25 25 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Born: 1935 In: Tupelo... Categories: Rock&Roll, American Singers, Academy Award winners...

26 26 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Born: 1935 In: Tupelo... ElvisPresley Categories: Rock&Roll, American Singers, Academy Award winners...

27 27 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Categories: Rock&Roll, American Singers, Academy Award winners... Born: 1935 In: Tupelo... born 1935 ElvisPresley bornin Tupelo

28 28 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Categories: Rock&Roll, American Singers, Academy Award winners... Born: 1935 In: Tupelo... AmericanSinger type ElvisPresley born bornin 1935 Tupelo

29 29 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Categories: Rock&Roll, American Singers, Academy Award winners... Born: 1935 In: Tupelo... AmericanSinger type ElvisPresley born bornin 1935 Tupelo won AcAward

30 30 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Categories: Rock&Roll, American Singers, Academy Award winners... Born: 1935 In: Tupelo... AmericanSinger type ElvisPresley born bornin 1935 Tupelo won locatedin AcAward USA

31 31 Extracting from Wikipedia Elvis Presley Elvis Presley was one of the best blah blah blub blah don t read this, listen to the speaker! blah blah blah blubl blah you are still reading this! blah blah blah blah blabbel blah Categories: Rock&Roll, American Singers, Academy Award winners... Born: 1935 In: Tupelo... Person? subclass Singer subclass AmericanSinger type ElvisPresley born bornin 1935 Tupelo won locatedin AcAward USA

32 Adding WordNet Person subclass Singer Scientist Person subclass Singer AmericanSinger subclass type born ElvisPresley bornin WordNet a lexicon of the English language developed at Princeton 1935 Tupelo won locatedin AcAward USA 32

33 33 Adding Time and Space Person subclass Singer LasVegas 1967 AmericanSinger subclass type PriscillaPresley married born ElvisPresley bornin We added a temporal and spatial dimension to facts and entities Tupelo won locatedin AcAward USA

34 34 YAGO: a large knowledge base YAGO is a knowledge base that combines WordNet classes and Wikipedia instances has time and space has a manually evaluated accuracy of 95% born Person subclass Singer AmericanSinger type ElvisPresley subclass bornin 1935 Tupelo won locatedin AcAward USA

35 35 YAGO: a large knowledge base Extractor Extractor Extractor GeoNames WordNet Extractor Extractor Extractor Extractor Try it online Try locally [WWW 2007, JWS 2008, WWW 2011 demo, AIJ 2013, WWW 2013 demo]

36 36 Example: YAGO about Elvis Try it online

37 37 YAGO: a large knowledge base 100m facts 10m entities 95% accuracy 100 Web site visitors/day 1000 citations linked to Freebase and DBpedia

38 38 SOFIE extracts by Reasoning Elvis died in 528 died 1955 born 1935

39 39 SOFIE extracts by Reasoning Elvis died in 528 died 1955 born 1935 occurs( Elvis, died in,528)

40 40 SOFIE extracts by Reasoning Elvis died in 528 Einstein died in 1955 died born occurs( Elvis, died in,528) occurs( Einstein, died in,1955)

41 41 SOFIE extracts by Reasoning Elvis died in 528 Einstein died in 1955 died born occurs( Elvis, died in,528) occurs( Einstein, died in,1955) died(einstein,1955), born(elvis, 1935)

42 42 SOFIE extracts by Reasoning Elvis died in 528 Einstein died in 1955 died born occurs( Elvis, died in,528) occurs( Einstein, died in,1955) died(einstein,1955), born(elvis, 1935) occurs(x,p,y) & means(x,x) & R(X,Y) => pattern(p,r) occurs(x,p,y) & means(x,x) & pattern(p,r) => R(X,Y)

43 SOFIE extracts by Reasoning Elvis died in 528 Einstein died in 1955 died born occurs( Elvis, died in,528) occurs( Einstein, died in,1955) died(einstein,1955), born(elvis, 1935) occurs(x,p,y) & means(x,x) & R(X,Y) => pattern(p,r) occurs(x,p,y) & means(x,x) & pattern(p,r) => R(X,Y) born(x,y) & died(x,z) => Z>Y... 43

44 44 SOFIE extracts by Reasoning Elvis died in 528 Solving a Weighted MAX SAT problem at scale occurs( Elvis, died in,528) occurs( Einstein, died in,1955) died(einstein,1955), born(elvis, 1935) occurs(x,p,y) & means(x,x) & R(X,Y) => pattern(p,r) occurs(x,p,y) & means(x,x) & pattern(p,r) => R(X,Y) born(x,y) & died(x,z) => Z>Y...

45 45 SOFIE extracts by Reasoning Elvis died in 528 died 528 [WWW 2009] occurs( Elvis, died in,528) occurs( Einstein, died in,1955) died(einstein,1955), born(elvis, 1935) occurs(x,p,y) & means(x,x) & R(X,Y) => pattern(p,r) occurs(x,p,y) & means(x,x) & pattern(p,r) => R(X,Y) born(x,y) & died(x,z) => Z>Y...

46 Product extraction 46

47 Work on Ontologies Constructing ontologies Singer Applying ontologies Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 47

48 48 Rule Mining finds patterns married haschild haschild married(x, y) haschild(x, z) haschild(y, z)

49 Rule Mining finds patterns married haschild haschild married(x, y) haschild(x, z) haschild(y, z) But: Rule mining needs counter examples and RDF ontologies are positive only 49

50 50 Partial Completeness Assumption haschild Assumption: If we know r(x,y1),..., r(x,yn), then all other r(x,z) are false.

51 51 Partial Completeness Assumption haschild Assumption: If we know r(x,y1),..., r(x,yn), then all other r(x,z) are false.

52 52 Partial Completeness Assumption haschild Assumption: If we know r(x,y1),..., r(x,yn), then all other r(x,z) are false.

53 53 Partial Completeness Assumption haschild Assumption: If we know r(x,y1),..., r(x,yn), then all other r(x,z) are false. + an efficient implementation

54 54 AMIE finds rules in ontologies AMIE (5min) type(x, pope) diedin(x, Rome)

55 55 AMIE finds rules in ontologies AMIE (5min) type(x, pope) diedin(x, Rome) [WWW 2013]

56 Work on Ontologies Constructing ontologies Singer Applying ontologies Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 56

57 Ontologies are complementary plays type Singer RockSinger type married born YAGO birthdate 1935 ElvisPedia 57

58 No Links => No Use plays type Singer Wo is the spouse of the guitar player? RockSinger type married born birthdate 1935 YAGO ElvisPedia 58

59 Match classes, entities, & relations plays type Singer subclassof sameas RockSinger type married born subpropertyof birthdate 1935 YAGO ElvisPedia 59

60 Match classes, entities, & relations Elvis name Elvis label

61 Match classes, entities, & relations Elvis name Elvis label 1. Match literals

62 Match classes, entities, & relations Elvis name Elvis label 1. Match literals

63 Match classes, entities, & relations Elvis name Elvis label 2. Assume small equivalence of all relations

64 Match classes, entities, & relations Elvis name Elvis label 2. Assume small equivalence of all relations

65 Match classes, entities, & relations Elvis name Elvis label 3. If entities share a relation that is highly inverse functional, and object is matched, match them.

66 Match classes, entities, & relations Elvis name Elvis label 3. If entities share a relation that is highly inverse functional, and object is matched, match them.

67 Match classes, entities, & relations Elvis name Elvis label 4. If relations share many pairs, increase their match

68 Match classes, entities, & relations Elvis name Elvis label 4. If relations share many pairs, increase their match

69 Match classes, entities, & relations Elvis name Elvis label 5. Iterate P (e 1 e 2 ) = 1 42 α β... P (r 1 r 2 )... P (r 1 r 2 ) = 42φ... P (e 1 = e 2 )...

70 Match classes, entities, & relations singer type type AmericanSinger 6. Compute class subsumption P (c 1 c 2 ) = arcsin(4.1125) P (e 1 e 2 )..

71 Match classes, entities, & relations singer type type AmericanSinger 6. Compute class subsumption

72 PARIS:match entities,classes,relations PARIS matches DBpedia & YAGO in 2 hours with 90% accuracy [VLDB 2012] 72

73 73 Matching heterogeneous KBs bornincountry USA bornincity Tupelo locatedin USA

74 74 1. Coalesce the KBs bornincountry USA bornincity Tupelo locatedin

75 75 2. Mine rules bornincountry USA bornincity Tupelo locatedin AMIE bornincity(x, y) locatedin(y, z) bornincountry(x, z) ROSA rule

76 76 ROSA rules match ontologies [AKBC 2013] bornincity(x, y) locatedin(y, z) bornincountry(x, z) ROSA rule

77 Work on Ontologies Constructing ontologies Singer Applying ontologies Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 77

78 78 Plagiarism People may steal from other ontologies without giving due credit. Most ontologies have licenses that require attribution. cc stolen YAGO EvilPedia

79 Additive Watermarking By adding a few fake facts to the source ontology, one can prove theft in the target ontology. stolen YAGO EvilPedia [WWW 2012 demo] 79

80 Subtractive Watermarking One can also prove theft by selectively removing facts from the source ontology. stolen YAGO EvilPedia [ISWC 2011] 80

81 Work on Ontologies Constructing ontologies Singer Applying ontologies Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 81

82 82 Mining Le Monde

83 83 Mining Le Monde

84 84 Mining Le Monde

85 85 Mining Le Monde time place entity 1967 USA

86 86 Mining Le Monde Average age of people mentioned

87 Mining Le Monde [AKBC 2013] 87

88 Work on Ontologies Constructing ontologies Singer Applying ontologies Protecting ontologies type type born? born? Mining ontologies A B C Aligning ontologies 88

89 YAGO can answer our question Find x such that x is called Elvis x is a singer x was born after

90 YAGO can answer our question Find x such that x is called Elvis x is a singer x was born after 1970 Elvis Crespo, singer, born

91 Thank you for your attention! Applying ontologies Constructing ontologies Singer Protecting ontologies type type born? born? Aligning Mining ontologies ontologies A B C Slides done with Powerline, my free SVG slide editor with LaTex support 91