百科和佛學知識圖譜構建技術介紹

漆桂林東南大學認知智能研究所Schedule

of

My

Talk百科知識圖譜構建技術佛學知識圖譜構建技術IntroductionofKnowledgeBasesWhatisknowledge?Facts,information,descriptions,orskillsAcquiredthroughexperienceoreducationbyperceiving,discovering,orlearningKnowledgebase:anorganizedrepositoryofknowledgeconsistingofconcepts,instances,relations(properties),facts,rulesetc.Isaprincipalpartofexpertsystems“thepowerofanAIprogramcametobeseenaslargelyinitsknowledgebase”EdwardFeigenbaum,1994ACMTuringAwardDevelopmentofKnowledgeBaseinRecentDecades1985199019952000(#$capitalCity#$France#$Paris)student

enrollee

person35millionarticlesin288differentlanguages…15thousandconcepts600millioninstances20billionfacts200520102012NELLGoogle

Knowledge

Graph

(KG)It

isanewgenerationofintelligentsearchtechnology,whichenablesyoutosearchforthings,notstringsFormal

definition:

a

knowledge

graph

is

a

knowledge

base

with

graph

structure,

where

the

nodes

are

instances

or

concepts,

and

edges

are

relations

between

themIt

is

a

special

semantic

networkIt

belongs

to

knowledge

engineering中興通訊上市公司非上市公司子公司中興康訊Acacia(IPO中)卓翼科技美國高通共進股份宇順電子美國博通供應商客戶競爭對手合作伙伴中國移動英特爾華為中國聯(lián)通大富科技華星創(chuàng)業(yè)盛路通信超聲電子ExampleKG

and

Semantic

Search

Go

deeper

and

broaderTechnologiesofKnowledgeBaseConstructionBaiduHudongZh-WikipediaKnowledge

Graph

(KG)ConstructionfromOnlineEncyclopediasWell-knownopenknowledgegraphssuchasDBpedia,YagoandZhishi.mearebuiltfromonlineencyclopedias.Technologies

ofencyclopedicknowledgegraphconstruction:DataextractionEntitymatchingTypeinferenceZhishi.meZhishi.me(http://zhishi.me)isthefirstefforttopublishlargescaleChinesesemanticdataandlinkthemtogetherasaChineseLinkingOpenData(CLOD).OverviewofZhishi.meCurrently,itconsistsofstructureddataextractedfromthreelargestChineseencyclopediasites:BaiduBaikeHudongBaikeChineseWikipediaItnow

has

over

10

milliondistinctinstancesand200millionRDFtriples,

and

can

be

accessed

by

online

API,

lookup

service

and

SPARQL

endpoint.LabelsAbstractsRedirectsImagesrdfs:labelzhishi:abstractrdfs:commentdbpedia:abstractzhishi:pageRedirectszhishi:thumbnailDataExtractionXingNiu,XinruoSun,HaofenWang,ShuRong,GuilinQi,YongYu:Zhishi.me-WeavingChineseLinkingOpenData.ISWC2011:205-220infoboxPropertieshttp://zhishi.me/[sourceName]/property/[propertyName]http://zhishi.me/baidubaike/property/中文名稱“南京”@zhDataExtractionInternalLinkszhishi:internalLinkzhishi:categoryskos:broaderDataExtractionEntityMatchingBaidu:北京Zh-Wiki:北京市EquivalententitiesEntityMatchingAutomaticallydiscoveringandrefiningdataset-specificmatchingrulesiniterationsDerivingtheserulesbyfindingthemostdiscriminativedatacharacteristicsforagivendatasourcepair,

e.g.(baidu:北京,Zh-wiki:北京市).From

Haofen

WangForeachpairofexistingmatchedinstances,theirproperty-valuepairsaremerged.ValuesProperty_1Property_2“大熊貓”baidu:標簽hudong:中文學名“Ailuropodamelanoleuca”baidu:拉丁學名hudong:二名法“白鰭豚”baidu:標簽hudong:中文學名“桂花”baidu:標簽hudong:中文學名………EntityMatchingFrom

Haofen

WangMatchingrule(frequentsetmining):baidu:xandhudong:xarematched,iff.valueOf(baidu:標簽)=valueOf(hudong:中文學名)andvalueOf(baidu:拉丁學名)=valueOf(hudong:二名法)andvalueOf(baidu:綱)=valueOf(hudong:綱)EntityMatchingFrom

Haofen

WangApplyingtheobtainedrule(s)ontheunlabeleddatatogeneratematches’candidates.Thecombinerisusedtocombineconfidencevaluesofamatch’scandidate.EntityMatchingFrom

Haofen

WangType

InferenceTypeinformationstatingthataninstanceisofacertaintype(e.g.Chinaisaninstanceofcountry)isanimportantcomponentofknowledgebasesGivenanapplication

scenario—QuestionAnswering.Question:WhoistheNobellaureateinliteratureofpeople’s

republicofChina?Answer:Moyan.Howtogettheanswer?

MoyanInstanceOf

Nobellaureateofpeople’srepublicofChinaTianxingWu,ShaoweiLing,GuilinQi,HaofenWang:MiningTypeInformationfromChineseOnlineEncyclopedias.JIST2014:213-229The4th

JointInternationalSemanticTechnologyConferenceApproach

InChineseonlineencyclopedias,wediscoverthatlotsoffine-grainedtypesexistincategoriesofarticlepages.

“TimBerners-Lee”hasseveralcategories:“Englishcomputerscientists”,“PeopleassociatedwithCERN”,“EnglishexpatriatesintheUnitedStates”,“LivingPeople”,“WorldWideWebConsortium”

The4th

JointInternationalSemanticTechnologyConferenceApproachInChineseonlineencyclopedias,wediscoverthatlotsoffine-grainedtypesexistincategoriesofarticlepages.

Givenanexample:

Giventhearticlepagesof“China”inBaiduBaike,Hudong

BaikeandChineseWikipedia,itscategoriesareasfollows:

The4th

JointInternationalSemanticTechnologyConferenceApproach(cont.)Wetakethecategoriesofonegiveninstanceasitscandidatetypesandtrytofilteroutthenoiseleveragingtheattributes.

The4th

JointInternationalSemanticTechnologyConferenceApproach(cont.)Wetakethecategoriesofonegiveninstanceasitscandidatetypesandtrytofilteroutthenoiseleveragingtheattributes.

Intuitively,whengivenattributesofacertaininstanceasfollows:

“actors,releasedate,director”

aninstanceof“Movie”

“name,foreignname”aninstanceof“?”The4th

JointInternationalSemanticTechnologyConferenceApproach(cont.)Wetakethecategoriesofonegiveninstanceasitscandidatetypesandtrytofilteroutthenoiseleveragingtheattributes.

Intuitively,whengivenattributesofacertaininstanceasfollows:

“actors,releasedate,director”

aninstanceof“Movie”

“name,foreignname”aninstanceof“?”Weassumethatifaninstancecontainstherepresentativeattributesofonecandidatetype,theinstanceprobablybelongstothistype.

The4th

JointInternationalSemanticTechnologyConferenceApproach(cont.)Wetakethecategoriesofonegiveninstanceasitscandidatetypesandtrytofilteroutthenoiseleveragingtheattributes.

Intuitively,whengivenattributesofacertaininstanceasfollows:

“actors,releasedate,director”

aninstanceof“Movie”

“name,foreignname”aninstanceof“?”Weassumethatifaninstancecontainstherepresentativeattributesofonecandidatetype,theinstanceprobablybelongstothistype.

Butanotherproblemis:categoryattributesarenotabundantlyavailable.

The4th

JointInternationalSemanticTechnologyConferenceApproach(cont.)ExplicitIsARelationDetector:DetectexplicitinstanceOfandsubclassOfrelationsCategoryAttributesGenerator:

GenerateattributesforcategorieswithanattributepropagationalgorithmInstanceTypeRanker:

Rankcandidatetypeswithagraph-basedrandomwalkmethod

The4th

JointInternationalSemanticTechnologyConferenceExplicitIsARelationDetectorExplicitInstanceOfRelationDetection

MiningexplicitinstanceOfrelationfrominfoboxes

I={i1,i2,…,in}aninstancesetallarticlesallarticlecategoriesC={c1,c2,…,cm}aconceptsetAttributevaluea1v1a2v2……infobox{<a1,v1>…,<ak,vk>}anAVPsetThe4th

JointInternationalSemanticTechnologyConferenceExplicitIsARelationDetectorExplicitInstanceOfRelationDetection

MiningexplicitinstanceOfrelationfrominfoboxes

I={i1,i2,…,in}aninstancesetallarticlesallarticlecategoriesC={c1,c2,…,cm}aconceptsetAttributevaluea1v1a2v2……infobox{<a1,v1>…,<ak,vk>}anAVPsetvkinstanceOfakExample:<director,StevenSpielberg>

The4th

JointInternationalSemanticTechnologyConferenceExplicitIsARelationDetectorExplicitInstanceOfRelationDetection

MiningexplicitinstanceOfrelationfromabstracts

performdependencyparsingwithFudanNLP[Qiuetal.,2013]SubjectPredicateObjectThe4th

JointInternationalSemanticTechnologyConferenceExplicitIsARelationDetectorExplicitInstanceOfRelationDetection

MiningexplicitinstanceOfrelationfromabstracts

performdependencyparsingwithFudanNLP[Qiuetal.,2013]SubjectPredicateObject

The4th

JointInternationalSemanticTechnologyConferenceExplicitIsARelationDetectorExplicitInstanceOfRelationDetection

MiningexplicitinstanceOfrelationfromabstracts

performdependencyparsingwithFudanNLP[Qiuetal.,2013]SubjectPredicateObject邁克爾·喬丹instanceOf籃球運動員MichaelJeffreyJordanBasketballPlayer

The4th

JointInternationalSemanticTechnologyConferenceExplicitIsARelationDetectorExplicitSubclassOfRelationDetection

GeneratecandidateSubclassOfcategorypairsintheformof(sub-category,category)basedonthecategorysystem.Checkwhetherthe(sub-category,category)sharethesamelexicalhead

withPOStagging.

Foreach(sub-category,category),checkwhetherthecategoryisaparentconceptofthesub-categoryinZhishi.schema[Wangetal.,2014]江蘇學校(schoolinJiangSu)subclassOf中國學校(schoolinChina)The4th

JointInternationalSemanticTechnologyConferenceCategoryAttributesGeneratorWetakeattributesininfoboxtemplatesasexistingcategoryattributes

andattributesininfoboxofarticlepagesasinstanceattributes.

WeconstructaCategoryGraphcomposedofallcategorieswithsubclassOfrelations.

WepropagateattributesovertheCategoryGraphleveragingexistingcategoryattributes,instanceattributes,identifiedinstanceOfandsubclassOfrelations.The4th

JointInternationalSemanticTechnologyConferenceCategoryAttributesGeneratorTheattributepropagationalgorithmarebasedonfollowingrules:Rule1:Ifacategorychasattributesfrominfoboxtemplates,theseattributesshouldremainunchanged.Rule2:Ifacategorychassomeinstanceswithattributes,theattributesshouldbepropagatedtocwhentheyaresharedbymorethanhalfoftheseinstances.Rule3:Ifacategorychassomechildcategorieswithattributes,theattributesshouldbepropagatedtocwhentheyaresharedbymorethanhalfofthesechildcategories.Rule4:Ifparentcategoriesofacategorychaveattributes,alltheattributesshouldbeinheritedbyc.The4th

JointInternationalSemanticTechnologyConferenceInstanceTypeRankerWeorganizeeachgiveninstance,itsattributesandcategories(i.e.candidatetypes)ofthecorrespondingarticlepageintoanInstanceGraph.WegroupsynonymousattributeswithBabelNetbeforeconstructingallInstanceGraphs.The4th

JointInternationalSemanticTechnologyConferenceInstanceTypeRankerThe4th

JointInternationalSemanticTechnologyConferenceInstanceTypeRankerWeassumethatthefewercategoriesanattributebelongsto,themorerepresentativetheattributeis.The4th

JointInternationalSemanticTechnologyConferenceInstanceTypeRanker

Whenexecutingarandomstepfromthegiveninstancetooneofitsattributes,thewalktendstochoosethemostrepresentativeattributeinordertowalktothecorrectcategories.Whenexecutingarandomstepfromanattributetotheoneofthecategoriesinthearticlepage,thecategoriescontainingthisattributehaveequalopportunity.The4th

JointInternationalSemanticTechnologyConferenceExperimentAccuracyEvaluationWerandomlyselect500(category,attribute)pairsfromeachonlineencyclopediaand500typestatementsfromdifferentsourcesineachonlineencyclopedias.Weinvitesixpostgraduatestudentswhoarefamiliarwithlinkeddatatolabeltheeachsamplementionedabovewith"Correct","Incorrect",or"Unknown".Togeneralizefindingsoneachsampletothewholedataset,wecomputetheWilsonintervals[Brownetal.,2001]for=5%.The4th

JointInternationalSemanticTechnologyConferenceExperimentAccuracyEvaluationWerandomlyselect500(category,attribute)pairsfromeachonlineencyclopediaand500typestatementsfromdifferentsourcesineachonlineencyclopedias.Weinvitesixpostgraduatestudentswhoarefamiliarwithlinkeddatatolabeltheeachsamplementionedabovewith"Correct","Incorrect",orUnknown".Togeneralizefindingsoneachsampletothewholedataset,wecomputetheWilsonintervals[Brownetal.,2001]for=5%.The4th

JointInternationalSemanticTechnologyConferenceExperimentComparisonwithOtherKnowledgeBases(OverlapofTypeinformation)WecomparealltheobtainedChinesetypeinformationwiththatofotherwell-knownknowledgebases,namelyDBpedia,YagoandBabelNet.SinceDBpediaandYagohavemultilingualversions,wemappedtheEnglishtypestatementstoChineseones(bothinstanceandtypeinonetypestatementcanbemappedtotheChineselabels)The4th

JointInternationalSemanticTechnologyConferenceExperimentComparisonwithOtherKnowledgeBases(OverlapofTypeinformation)WecomparealltheobtainedChinesetypeinformationwiththatofotherwell-knownknowledgebases,namelyDBpedia,YagoandBabelNet.SinceDBpediaandYagohavemultilingualversions,wemappedtheEnglishtypestatementstoChineseones(bothinstanceandtypeinonetypestatementcanbemappedtotheChineselabels)TechnologiesofKnowledgeBaseConstructionWebAccesstoZhishi.me

http://zhishi.me/apiSchedule

of

My

Talk百科知識圖譜構建佛學知識圖譜構建Framework(takeBuddhistfiguresastheexample)KnowledgeCollectionCategory方法人工觀察百科中與佛教人物相關的分類抽取佛教人物分類下所有文章對應的實體命名規(guī)則方法

例：

“.+菩薩”“.+禪師”維基百科“佛教頭銜”分類下的所有實體已抽取出的實體名中高頻的公共字符串KnowledgeFusion主語融合實體的“別名”屬性和重定向作為實體的別名集合不同來源的實體存在一個完全匹配的別名則認為是相同實體人工檢查相同實體數(shù)多于三個的映射百度百科：互動百科：維基百科：{確吉堅贊，班禪額爾德尼·確吉堅贊，羅桑赤烈倫珠}{班禪額爾德