semantic technologies for data access

SemanticTechnologiesforDataAccess

DiegoCalvaneseFreeUniversityofBozen-Bolzano,Italy

[email protected]

MartinRezkRakuten Inc.

[email protected]

Rakuten TechnologyConference(RTC)22October2016,Tokyo 1

Outlineofthepresentation

1. IntroductiontoSemanticTechnologiesfordatamanagement(Diego)

2. UsecasesatRakuten (Martin)3. Ontology-baseddataaccesswithOntop (Diego)

2

1.IntroductiontoSemanticTechnologiesfordatamanagement

3

Semantics• Isabranchoflinguisticsconcernedwiththestudyofmeaning ofexpressionsinarepresentationscheme(i.e.,alanguage)

• Wenaturallyassociatemeaningtoexpressionsinnaturallanguage,e.g.,towordsandsentences.

• Semanticsneededalsoforartificialandformallanguages.

• Semanticsiscrucialwhenmachinesneedtointeractandexchangeinformationwithhumansandwitheachother.

4

MarsClimate Orbiter had some“misunderstanding”withground stationonimperial vs.metric units [1999].

5

327MUS$burned intheMarsatmosphere!

Implicitvs.explicitsemanticsofdata

id surname name #courses258 Lenzerini Maurizio 2262 Carlucci Gigina 1484 Nardi Daniele 3271 Catarci Tiziana 0435 Marchetti Alberto 2… … … …

6

FacultyStaffWhoisthedean?WhoworksinthedepartmentofNardi?

Thedeanistheoneteachingnocourses.Thedept isencodedinthefirstdigitoftheid.

Thishindersunderstanding,henceusability,maintenance,reusability,extensibility,…

Semanticsisrepresentedimplicitly!

Instead,wewanttorepresentsemanticsexplicitly!

Representingsemanticsexplicitly

• WeareusinghereanEntity-Relationshipdiagram,whichadmitsagraphicalnotation.

• Wecouldalsohaverepresentedthesemanticsusingaformalizationinlogic.

7

Staff Courseteaches

Department

belongsTo

surnamenameid

is_dean

BringingSemanticstoWebdata

“TheSemanticWebis[…]anextensionofthecurrentwebinwhichinformationisgivenwell-definedmeaning,betterenablingcomputersandpeopletoworkincooperation.”[T.Berners-Lee,J.Hendler,O.Lassila,2001]

8

SemanticWebTechnologies

• AretechnologiesenablingtheSemanticWebvisiontobecomereality.

• Basedonflexibledatarepresentationformats.

• Semanticsisrepresentedexplicitlybymeansofformalorlogic-basedlanguages.

9

Semantic Weblayers [2001]

TechnologiesforrepresentingWebdata

Weneedmechanismsforrepresentingdatainaflexibleway:• XML (ExtensibleMarkupLanguage): markuplanguageforencodingdocumentsinahumanandmachinereadableformat(bytheW3C).

• JSON: (JavaScriptObjectNotation): lightweightdata-interchangeformat(byEcma International).

• RDF (ResourceDescriptionFramework): flexibledatamodelbasedontheideaofmaking statements about (web)resources intheformofsubject–predicate–objecttriples(bytheW3C).

10

Aflexibledataformatisnotsufficient

• XML,JSON,RDFdefinemechanismstoassertfactsaboutdataitems:tiziana rdf:type Staff .databases rdf:type Course .tiziana teaches databases.

• Buttheydonotprovidemeanstovalidatethedata:csEngineering rdf:type Department .dumbo rdf:type Animal .dumbo teaches csEngineering .

11

Addingsemanticstodata

TheW3Chasdefinedseverallanguagesforrepresentingthesemanticsexplicitlyandforinterlinkingdata:

• RDFS (RDFSchema): lightweightschemalanguagefordescribingconceptsandtheirrelationships.

• OWL (WebOntologyLanguage): veryexpressiveontologylanguageformodelingknowledgeaboutadomainofinterest.

• RIF (RuleInterchangeFormat): rule-basedlanguageforthespecificationofknowledge.

• StandardsforLOD (LinkedOpenData)12

LinkedOpenData(LOD)

• Microdata

• RDFa

• JSON-LD

• Microformats

13

SeveralstandardshavebeendefinedtoannotateandinterlinkwebpageswithHTML-embeddeddata:

SeeinvitedtalkbyChrisBizer atISWC2016(Kobe)http://www.slideshare.net/bizer/is-the-semantic-web-what-we-expected-adoption-patterns-and-contentdriven-challenges-iswc-2016-keynoteWebDataCommons(11/2015):HTML-embeddeddataisprovidedby• 19%oftheprimary-level-domains(2.72Moutof14.41M)• 30%oftheHTMLpages(540Moutof1.71B)

Example ofRDFSspecification

belongsTo rdfs:domain Staff .belongsTo rdfs:range Department .teaches rdfs:domain Staff .teaches rdfs:range Course .GradCourse rdfs:subClassOf Course .…

Essentially,inRDFSwe canexpressthesame informationas inanERdiagram(except formultiplicities andcomplete/disjoint hierarchies).

Wecanencodethesemanticsofthedomain,anduseittocheckwhetherstatementsaremeaningful.

14

Staff Courseteaches

Department

BelongsTo

surnamename

id

is_dean

GradCourse

Someusecases forsemantictechnologies

1. Toprovideuserswithnewdiscoveryaxes(Rakuten Ichiba,PriceMinister).

2. Todomorefine-grainedpersonalizationinmarketingcampaigns(Rakuten Ichiba).

3. Foraccessingandintegratingheterogeneousdatasources(e.g.,atStatoilandSiemens).

15

2.Semantic technologies atRakuten

16

GivingSemanticstoData

MartinRezkRakuten [email protected]

(JointworkwithBrunoCharron,Hirate Yu,andDavidPurcell)Publishedintheproc.ofISWC’16

Rakuten TechnologyConferenceOctober22sd,2016

17

• Ichiba offersaround200Mitemsclassifiedinalargelegacy taxonomy(~40,000classes)

• Eachoftheseitemshasapagedescriptioncreatedbythemerchants.

R a k u t e n G r o u p

18

Why semant i c s?

19

What bus iness need t r i g ge red th i s p ro jec t

MarketingTeam( D a t a A n a l y s i s )

20

21

material

shape

size

color

maker

origin W H AT P R O P E R T I E S A R E I M P O R TA N T I N E A C H C AT E G O R Y ?

period

22

Poireround

Demi-lune

rectangularoval

WHAT VA LU E S E X I S T ?

material

shape

size

color

maker

origin

period

23

BU T WH I C H H A S T H E MO S T GMS ?

Poireround

Demi-lune

rectangularoval

material

shape

size

color

maker

origin

period

24

BU T WH I C H I S T R END I NG ?

material

shape

size

color

maker

origin

period

Poireround

Demi-lune

rectangular

oval

CatalogTeam( C omp . - U s e r I n t e r a c t i o n )

Whataretherelevantdiscoveryaxesthatcanhelptheuserstoexplorecomplexclasses?

What bus iness need t r i g ge red th i s p ro jec t?( Fo c u s i n g o n I c h i b a fo r n ow )

25

CatalogTeam( C omp . - U s e r I n t e r a c t i o n )

Whataretherelevantdiscoveryaxesthatcanhelptheuserstoexplorecomplexclasses?

What bus iness need t r i g ge red th i s p ro jec t?

26

So lu t ion : Ex t rac t ing semant i c i n fo rmat ion

G i v i n g S e m a n t i c s t o

D a t a

• Wewanttoextractrelevant datapropertiesforcomplexandprofitablesubclasses.

• Foreachdataproperty,wewanttoextracttherelevantsubsetofitsrange.

• Wedonotwanttomodelthewholedomain.

• Wewanttolinkthepropertiesandvaluesbacktotheitems.

• Wewantthesolutiontobeasmuchaspossiblelanguageindependent.

27

Overv iew o f the approach

28

C L A S S S E L E C T I O N

29

<html><tr><tc></tc></tr>

</html>


</html>

SeedGeneration

C L A S S S E L E C T I O N P R O P E R T Y E X T R A C T I O N

U s i n g s t a n d a r d d a t a m i n i n g t e c h n i q u e s a n d n o v e l m a t h e m a t i c a l m o d e l s t o c l e a n t h e l i s t .

30


</html>


</html>

小倉百人一首の雅なおせんべい！高級感のあるパッケージで、ちょっとずつ

食べようと思っても、ついつい手が出てしまうので、隠しつつちょこちょこ他寝るように気をつけています。おいしすぎも、困りものですよ・・・

おせんべい★お試しセットもち吉さんのおせんべいが食べたくなって、買い

に行かなくても済むんで宅配たすかります～味は変わらず美味しいですよ♪

気心の知れた方に差し上げても、たいへんに喜ばれますね▼・∀・▼ありがとう

SeedGeneration

Boot Strapping


We u s e m a c h i n e l e a r n i n g t o e x t e n d t h e o r i g i n a l s e e d . 31


</html>


</html>






SeedGeneration

Bootstrapping

C L A S S S E L E C T I O N P R O P E R T Y E X T R A C T I O N L I N K I N G

We g e n e r a t e t r i p l e s o f t h e f o r m :( I t e m 1 8 5 , O r i g i n , J a p a n ) 32

Subt ree Ex t rac t ion

33

C L A S S S U B T R E E S E L E C T I O N

We m e a s u r e :- H o m o g e n e i t y ( f o r t h e u s e r s )- N e e d o f N a v i g a t i o n a l A s s i s t a n c e ( i n t e r e s t i n g t o t h e

u s e r ? )

W H E R E W E S TA R T : T H E TA X O N OM Y

S e l e c t i n g w h i c h c l a s s e s a m o n g t h e 4 0 . 0 0 0 w e w i l l w o r k w i t h .

34

H OM O G E N E I T Y

T H E TA X O N OM Y T R E E W E S E E

S u b t r e e E x t r a c t i o n

T H E TA X O N OM Y T R E E T H E U S E R S S E E

35


T H E TA X O N OM Y T H E U S E R S S E E

S u b t r e e E x t r a c t i o n


36


A l c o h o lW i n e

R e dW h i t eR o s e

W i n e


37



R e dW h i t eR o s e U s u a l W i n e s( W h i t e a n d R e d )

F a n c yW i n e s( R o s e )

W i n e

Lookingatusers’shoppingbehavior

T H E TA X O N OM Y T R E E T H E U S E R S S E E


38



R e dW h i t eR o s e U s u a l W i n e s( W h i t e a n d R e d )

F a n c yW i n e s( R o s e )

W i n e

Weselectthe``right’’subtreestoextractpropertiesfrom

H o m o g e n o u s

T H E TA X O N OM Y T H E U S E R S S E E


39

S U B T R E E S E L E C T I O N

N e e d o f N a v i g a t i o n a l A s s i s t a n c e

- Givenasubtree,wecomputeitsGMSdiversity- 𝑒 #$% & ∗ ) * ( % & )

- ( exponentialoftheShannonentropy)

- Wherepi istheproportionofthetotalGMSofthesubtreewhichisduetotheitemi.

- Intuitively,itrepresentstheeffectivenumberofitemsinasubtreemakingupitsGMS.

~ 2 . 7

40

S U B T R E E S E L E C T I O N- Givenasubtree,wecomputeitsGMSdiversity

- 𝑒 #$% & ∗ ) * ( % & )

- ( exponentialoftheShannonentropy)

- Wherepi istheproportionofthetotalGMSofthesubtreewhichisduetotheitemi.

- Intuitively,itrepresentstheeffectivenumberofitemsinasubtreemakingupitsGMS.

- Asubtreeissaidtohaveahighneedfornavigationalassistance(NNA)ifitseffectivenumberofitemsismorethan215.

N e e d o f N a v i g a t i o n a l A s s i s t a n c e

41


</html>


</html>

SeedGeneration


42

Proper ty Va lue Ex t rac t ion(Seed )

43


</html>


</html>

SeedGeneration

P r o p e r t y Va l u e E x t r a c t i o n

Givenasubtreet,weextracttheinitialsetofpropertiesandvalues(PV)fromHTMLtablesandsemi-structuredtextinputbymerchantsint.

PropertyCandidates

PossibleValues 44


</html>


</html>

SeedGeneration

P r o p e r t y Va l u e E x t r a c t i o n

Thissetcontainssomeissues:

- Redundantpropertynames:e.g.Maker/Producer

- Noisypropertyvalues.

- UselessPVfordiscoveryaxes:e.g.expirationdate

45


</html>


</html>






SeedGeneration

Boot Strapping


46

Bootst rapp ing

47

• Word2vecisashallowwordembeddingmodel.

• Themodellearnstomapeachdiscretewordid(0throughthenumberofwordsinthevocabulary)intoalow-dimensionalcontinuousvector-space.

• Wordswithsimilardistributionalproperties(i.e.,thatco-occurregularly)tendtosharesomeaspectofsemanticmeaning.

B o o t s t r a p p i n g : w o r d 2 v e c

Words

Vectors

whiteblueturquoise

electronMozarttango

Similar

NotSimilar48

• Wetrain2differentmodelswithdifferentparameters(cbow,skipgram,window,etc.)andchunkingmethods.

• Weiteratingoverthepropertiesextendingtheknownrange…asfollows:

Shape


49



Shape

OvalRectangle


50



OvalRectangle

Moon

Round

Shape UsingthetwoWord2Vecmodels


51



OvalRectangle

Round

Shape


52



Material


53

B o o t s t r a p p i n g

Taketheseedvaluesandgetthemostsemanticallysimilarwordsineachmodel…

TaketheintersectionX…

Ifthewordisnotabettervalueforadifferentproperty

ForeachwordinX

Addittothepropertyvalues.

ForeachpropertyPi…

54

Resu l t s ( a t the t ime o f pub l i cat ion )

55

Evaluat ion(at the t ime of publ icat ion)

• Weevaluatedourresultsin4categories:rice,beef,wine,andnecktie.

• Riceandbeefhadbeenpreviouslyextendedbythecatalogteam.

• TheresultsfromwineandnecktiewereevaluatedbyRakuten members.

56

Humancomparison- beef

Totalproperties Intersection Difference TotalValuesValues forIntersectedproperties

Values nonintersectedproperties

Manually Extracted 5 3 2 49 18 31

AutomaticallyExtracted 9 3 6 118 43 75

Cut

Type

Locality

Size

Intended Use

Ingredient

Allergens

Shippingfee

Processingarea

Countryofproduction

ProductName

Resu l t s Beef :

57

Subtree Count Overall Max Median Mean Min

Rice 6 0.92 1.00 0.97 0.81 0.20

Beef 9 0.86 1.00 0.88 0.83 0.50

Wine 9 0.91 1.00 0.81 0.77 0.20

Necktie 8 0.88 1.00 0.85 0.70 0.00

Theabovetableshoesthenumberofproperties,theoverallaccuracyofthepropertiesandthedistributionoftheaccuraciesbyproperty

Accuracy :

58

Conc lus ions

59

60

ConclusionsT h e B e a u t y T h e B e a s t T h e K n i g h t

Rakuten canimproveallitsservicesbyextractingsemanticinformation.

Thereistoomuchdata,spreadacrossdatasources,andthesemanticsishiddeninusersshoppinglogs,databases,andtext.

Withourapproach,wecandiscoverthesemanticshiddeninthedata,andbringittolighttobeused.

3.Ontology-Based DataAccesswithOntop

61

Howmuchtimeisspentsearchingfordata?

62

Engineersinindustryspendasignificantamountoftheirtimesearchingfordatathattheyrequirefortheircoretasks.Forexample,intheoil&gas industry,30–70%ofengineers’timeisspentlookingfordataandassessingitsquality.[Crompton,2008]

StatoilExploration

63

Facts:• 1,000TBofrelationaldata• usingdiverseschemata• spreadover2,000tables,overmultipleindividualdatabases

DataAccessforExploration:• 900expertsinStatoilExploration.

Expertsingeologyandgeophysicsdevelopstratigraphicmodelsofunexploredareasonthebasisofdataacquiredfrompreviousoperationsatnearbylocations.

Howmuchtime/moneyisspentsearchingfordata?

64

AuserqueryatStatoilShowallnorwegian wellboreswithsomeaditional attributes(wellboreid,completiondate,oldestpenetratedage,result).Limittoallwellboreswithacoreandshowattributeslike(wellboreid,corenumber,topcoredepth,basecoredepth,intersectingstratigraphy).LimittoallwellboreswithcoreinBrentgruppen andshowkeyatributes inatable.AfterconnectingtoEPDS(slegge)wecouldforinstancelimitfuther tocoresinBrentwithmeasuredpermeabilityandwhereitislargerthanagivenvalue,forinstance1mD. WecouldalsofindoutwhethertherearecoresinBrentwhicharenotstoredinEPDS(basedonNPDinfo)andwheretherecouldbepermeabilityvalues.Someofthemissingdatawepossiblyown,othernot.

SELECT [...]FROMdb_name.table1 table1,db_name.table2 table2a,db_name.table2 table2b,db_name.table3 table3a,db_name.table3 table3b,db_name.table3 table3c,db_name.table3 table3d,db_name.table4 table4a,db_name.table4 table4b,db_name.table4 table4c,db_name.table4 table4d,db_name.table4 table4e,db_name.table4 table4f,db_name.table5 table5a,db_name.table5 table5b,db_name.table6 table6a,db_name.table6 table6b,db_name.table7 table7a,db_name.table7 table7b,db_name.table8 table8,db_name.table9 table9,db_name.table10 table10a,db_name.table10 table10b,db_name.table10 table10c,db_name.table11 table11,db_name.table12 table12,db_name.table13 table13,db_name.table14 table14,db_name.table15 table15,db_name.table16 table16WHERE [...]

table2a.attr1=‘keyword’ ANDtable3a.attr2=table10c.attr1 ANDtable3a.attr6=table6a.attr3 ANDtable3a.attr9=‘keyword’ ANDtable4a.attr10 IN (‘keyword’) ANDtable4a.attr1 IN (‘keyword’) ANDtable5a.kinds=table4a.attr13 ANDtable5b.kinds=table4c.attr74 ANDtable5b.name=‘keyword’ AND(table6a.attr19=table10c.attr17 OR(table6a.attr2 IS NULL ANDtable10c.attr4 IS NULL)) ANDtable6a.attr14=table5b.attr14 ANDtable6a.attr2=‘keyword’ AND(table6b.attr14=table10c.attr8 OR(table6b.attr4 IS NULL ANDtable10c.attr7 IS NULL)) ANDtable6b.attr19=table5a.attr55 ANDtable6b.attr2=‘keyword’ ANDtable7a.attr19=table2b.attr19 ANDtable7a.attr17=table15.attr19 ANDtable4b.attr11=‘keyword’ ANDtable8.attr19=table7a.attr80 ANDtable8.attr19=table13.attr20 ANDtable8.attr4=‘keyword’ ANDtable9.attr10=table16.attr11 ANDtable3b.attr19=table10c.attr18 ANDtable3b.attr22=table12.attr63 ANDtable3b.attr66=‘keyword’ ANDtable10a.attr54=table7a.attr8 ANDtable10a.attr70=table10c.attr10 ANDtable10a.attr16=table4d.attr11 ANDtable4c.attr99=‘keyword’ ANDtable4c.attr1=‘keyword’ AND

table11.attr10=table5a.attr10 ANDtable11.attr40=‘keyword’ ANDtable11.attr50=‘keyword’ ANDtable2b.attr1=table1.attr8 ANDtable2b.attr9 IN (‘keyword’) ANDtable2b.attr2 LIKE ‘keyword’% ANDtable12.attr9 IN (‘keyword’) ANDtable7b.attr1=table2a.attr10 ANDtable3c.attr13=table10c.attr1 ANDtable3c.attr10=table6b.attr20 ANDtable3c.attr13=‘keyword’ ANDtable10b.attr16=table10a.attr7 ANDtable10b.attr11=table7b.attr8 ANDtable10b.attr13=table4b.attr89 ANDtable13.attr1=table2b.attr10 ANDtable13.attr20=’‘keyword’’ ANDtable13.attr15=‘keyword’ ANDtable3d.attr49=table12.attr18 ANDtable3d.attr18=table10c.attr11 ANDtable3d.attr14=‘keyword’ ANDtable4d.attr17 IN (‘keyword’) ANDtable4d.attr19 IN (‘keyword’) ANDtable16.attr28=table11.attr56 ANDtable16.attr16=table10b.attr78 ANDtable16.attr5=table14.attr56 ANDtable4e.attr34 IN (‘keyword’) ANDtable4e.attr48 IN (‘keyword’) ANDtable4f.attr89=table5b.attr7 ANDtable4f.attr45 IN (‘keyword’) ANDtable4f.attr1=‘keyword’ ANDtable10c.attr2=table4e.attr19 AND(table10c.attr78=table12.attr56 OR(table10c.attr55 IS NULL ANDtable12.attr17 IS NULL))

AtStatoil,ittakesupto4daystoformulateaqueryinSQL.

Statoillosesupto50M€peryearbecauseofthis.

Needforabstraction

Weneedtofacilitateaccesstodata:• byabstractingawayfromhowthedataisstored,and• bymakinguseofahighlevelviewonthedata,throughanontology.

65

Ontology-BasedDataAccess(OBDA)framework

66

• Isaplatformtoquerydatabases throughontologies,relyingonsemantictechnologies.

• CompliantwiththestandardsoftheW3C.• SupportsallmajorrelationalDBsvai JDBC(Oracle,DB2,MSSQLServer,Postgres,MySQL,H2,…).

• Open-source andreleasedunderApachelicense.• DevelopmentofOntop:

• developmentstarted6yearsago• alreadywellestablished:

• 200membersinthemailinglist• +7000downloadsinlast18months

• maindevelopmentcarriedoutinthecontextoftheEUprojectOptique 67

http://ontop.inf.unibz.it

Queryansweringbyrewriting

68

Ontological query

Rewritten query

SQLqueryRelational answer

Ontological answer Rewriting

Unfolding

Result Translation

Evaluation

Thankyouforlistening

69

FRAZZ:©JeffMallett/Dist.byUnited Feature Syndicate,Inc.

Any questions?DiegoCalvanese – [email protected] – [email protected]

semantic technologies for data access

Data & Analytics