Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
1
ColdStartKnowledgeBasePopulationatTAC2017TaskDescription1
Version1.0ofMay23,2017
What’sNew.............................................................................................................................................2Introduction...........................................................................................................................................3Input..........................................................................................................................................................5Schema................................................................................................................................................................5DocumentCollection......................................................................................................................................6EvaluationQueries.........................................................................................................................................8
ColdStartKBTaskOutput..............................................................................................................11Nodes.................................................................................................................................................................12Predicates........................................................................................................................................................13type.....................................................................................................................................................................................13*mentionpredicates...................................................................................................................................................14link......................................................................................................................................................................................16SF,sentiment,andeventpredicates....................................................................................................................16
EventRealis....................................................................................................................................................16Provenance.....................................................................................................................................................17ConfidenceMeasure.....................................................................................................................................18Comments........................................................................................................................................................19
SlotFillingTaskOutput...................................................................................................................19Evaluation............................................................................................................................................21ComponentEvaluations..............................................................................................................................21CompositeEvaluationAssessment.........................................................................................................22CompositeEvaluationScoring..................................................................................................................24
Submissions.........................................................................................................................................26Appendix..............................................................................................................................................27..................................................................................................................................................................28ChangeHistory...................................................................................................................................36
1TheTACorganizingcommitteewelcomescommentsonthisTaskDescription,oronanyaspectoftheTACevaluation.Pleasesendcommentstotac-kbp@nist.gov.
2
What’sNewThe2017ColdStartKBandSFtasksdifferfromthe2016tasksinthefollowingsignificantways:
1. TheColdStartKBincludesevents(analogoustoentities)andeventarguments(analogoustoslotfillers).
2. TheColdStartKBincludessentimentfromanentitytowardsanotherentity.3. TheColdStartKBwillbeevaluatedviaacompositeKBevaluationusingqueries(asinCold
Start2016),andasetofcomponentevaluationsforEntityDiscoveryandLinking(EDL),SlotFilling(SF),EventNuggetDetectionandCoreference(EN),EventArgumentandLinking(EAL),andSentiment.
4. MeanAveragePrecision(MAP)willbethemainevaluationmetricforthecompositeKBevaluationandthecomponentSFevaluation.
5. MultiplejustificationsareallowedandencouragedforKBrelations(involvingSF,event,orsentimentpredicates).JustificationspansforanysinglejustificationintheKBandSFtasksmustcomefromthesamedocument,inordertomakejustificationseasiertodifferentiateandcount.
The2017ColdStartKBConstructiontaskbuildsonthe2016taskbyextendingtheKBschematoincludenotonlyentitiesandSlotFillingrelations,butalsoeventsandrelationsinvolvingeventargumentsandsentimentbetweenentities.
Inaddition,ColdStart2017hasthegoalofencouragingsystemsto1)providemeaningfulconfidencevaluesforassertionsthataremadeintheKBand2)returnasmuchevidenceforeachKBrelationascanbefoundinthedocumentcollection.Toencouragedevelopmentofmeaningfulconfidencevalues,theprimaryevaluationmetricforColdStart2017willbeavariantofmeanaverageprecision(MAP).TosupportanevaluationthatrewardssystemsforfindingmorethanonejustificationforeachKBrelation,ColdStart2017requiresthatalltextspansforasinglejustificationmustcomefromasingledocument.ThissimplificationallowsTACtodefinetwojustificationstobethesameifandonlyiftheycomefromthesamedocument,andgivesmorecreditforfindingmoredocumentsthateachjustifyrelation.2
ThesubmittedColdStartKBsareevaluatedbybothacompositequery-basedevaluation,andasetofcomponentevaluations.ThecompositeKBevaluationappliesasetofColdStartevaluationqueriestoeachKBandassessesthecorrectnessoftheevents,sentimentsourcesandtargets,andSFslotfillersfound.
ThecomponentevaluationsareimplementedbyprojectingouttheindividualcomponentsfromthesubmittedKBandevaluatingeachcomponentoutputfileasthoughithadbeensubmitteddirectlytothestandalonetrackforthatcomponent.
2TheTACColdStartKBandSFtasksin2014-2016allowedasinglerelationjustificationtocontainprovenancespansfrommultipledocuments,inordertoencourageinferenceacrosswidercontexts.WhilerestrictingprovenancespanstoasingledocumentperjustificationmayseemlikeastepbackwardsforKBP2017,thebenefitofbeingabletocountjustificationsoutweighsthereductioninallowableinference.Furthermore,becauseofthetypesofrelationsandwhatisconsidercorrectinColdStart/SF,mostsuccessfulcross-documentinferenceinColdStarthasbeenlimitedtoahandfulofhard-codedinferencerules(e.g.,involvingfamilialrelationshipsandpart-wholerelationshipsbetweenGPE’s);suchinferencescouldbedonebysomedownstreamprocess,usingaseparateinferenceengineandapossiblyrichersetofinferencerulesandworldknowledge.
3
Thisdocumentdescribesthe2017composite(end-to-end)ColdStartKBConstructiontaskandthecomponentSlotFillingtask,whichareevaluatedusingpost-submissionassessmentofresponsestoColdStartevaluationqueries.
ThestandaloneEDL,EN,EAL,andsentimenttasksareevaluatedusinggoldstandardannotationsonacommonsetofapproximately500"core"documents,andaredescribedfullyontheirrespectivetrackhomepages.Thedetailedtaskdescriptionforthosecomponenttasksareavailableontheirtrackhomepages:
• EntityDiscoveryandLinking(http://nlp.cs.rpi.edu/kbp/2017/index.html).• EventNuggetDetectionandCoreference(http://cairo.lti.cs.cmu.edu/kbp/2017/event/)• EventArgumentandLinking(https://tac.nist.gov/2017/KBP/Event/Argument/)• BeliefandSentiment(http://www.cs.columbia.edu/~rambow/best-eval-2017/)
IntroductionSince2009,TACKBPhasevaluatedperformanceonseveralimportantaspectsofknowledgebasepopulation:entitydiscoveryandlinking,slotfilling,eventnuggetdetectionandcoreference,eventargumentextractionandlinking,andbeliefandsentiment.ThegoaloftheColdStarttrackistoexercisetechnologyineachoftheseareas,andevaluatetheabilityofasystemtousethesetechnologiestoactuallyconstructaknowledgebase(KB).ColdStartparticipantsbuildasoftwaresystemthatprocessesalargetextcollectionandcreatesaknowledgebasethatisconsistentwithandaccuratelyrepresentsthecontentofthatcollection.Theknowledgebaseisthenevaluatedasasingleconnectedresource,usingqueriesthattraversenodesandpredicatelinksintheKBtodetermineiftheKBcontainscorrectrelationsbetweenentities,events,andstrings.InColdStart2017,thepredicatescanbeslotfillingpredicates,sentimentpredicates,oreventargumentpredicates.
WecallthetaskColdStartKnowledgeBasePopulationtoconveytwofeaturesoftheevaluation:itimpliesboththataknowledgebaseschemahasbeenestablishedatthestartofthetask,andthattheknowledgebaseisinitiallyunpopulated.Thus,weassumethataschemaexistsfortheentities,events,andrelationsthatwillcomposetheknowledgebase;itisnotpartofthetasktoautomaticallyidentifyandnamerelationshipspresentinthetextcollection.In2017,ColdStartusesaschemathatcombinestheentitytypes,eventtypes,andrelationsfromtheTACcomponenttracksofEDL,SF,EN,EAL,andBeSt.
ColdStartalsoimpliesthattheknowledgebaseisinitiallyempty.ToavoidsolutionsthatrelyonverifyingcontentalreadypresentinWikipediaorotherlargedatasourcesaboutentities,thequeriesusedinColdStartwillbedominatedbyentitiesthatarenotpresentinWikipedia.
Thisdocumentdescribesboththeend-to-endColdStartKBconstructiontaskandthecomponentSlotFillingtask.
1. IntheColdStartKnowledgeBasetask(CSKB),participantssubmitentireknowledgebases,withoutpriorknowledgeoftheevaluationqueries.
2. IntheSlotFillingtask(SF),theColdStartevaluationqueriesthatinvolveonlySFpredicatesaresplitintoColdStartSlotFillingqueries,withoneentrypointperquery,andaredistributedatthestartofthetaskevaluationwindow.Participantsdonothavetosubmitentireknowledgebases.Rather,theyapplytheirslotfillingsystemtwice,thefirsttimeontheentrypointforeachquery,thesecondtimeoneachoftheresultsofthefirstround.
ParticipatingsystemsinboththeColdStartKBandSFtaskswillreceivethefollowinginput:
4
1. adocumentcollection;2. aknowledgebaseschema
Fromthese,ColdStartKBsystemswillproduceaknowledgebase.ThisKBwillbesubmittedtoNISTasasetofaugmentedsubject-predicate-objecttriples.TheColdStartKBwillincludevarious*mention, link,andtypetriples,aswellasarangeoftriplesinvolvingSF,eventargument,andsentimentpredicates(alltriplesaredescribedmorefullybelow).ParticipatingKBsystemsmusttieeachentitymentionandeventmentioninthedocumentcollectiontoaparticularKBnode;inthisway,theknowledgebasecanbequeriedwithoutfirstaligningittoareferenceknowledgebase.
SystemsparticipatingintheSlotFillingtaskwillalsoreceive:
3. asetofColdStartSlotFilling(CSSF)evaluationqueries(eachevaluationqueryisasequenceofoneortwoSFqueriestobeappliedinseries).
ForbothCSKBandSFtasks,theresultswillthenbeevaluatedbyNIST:
• SystemsparticipatingintheSlotFillingtaskreturnslotfillersdirectlyinresponsetothegivenCSSFevaluationqueries,andthefillersarethenassessedandscored.
• EvaluationoftheKnowledgeBasevariantwillstartbyapplyingtheColdStartevaluationqueriestothesubmittedknowledgebase.Eachquerywillstartatanamedentitymentioninadocument(identifiedbythequery’sandtags),identifytheknowledgebaseentitythatcorrespondstothatmention,followasequenceofoneormorerelationswithintheknowledgebase,andendina“slotfill”(whichcouldbeanentity,event,orstringnode,dependingontherelationpredicate).TheresultingslotfillswillbeassessedandscoredinthesamewayasintheSlotFillingvariant.Forexample,aCSSFevaluationquerymightask‘whataretheagesofthesiblingsoftheBartSimpson4mentionedinDocument42?’AsystemthatcorrectlyidentifieddescriptionsofBart’ssiblingsinthedocumentcollection,linkedthemtotheappropriatenodeintheKB,andalsofoundevidenceforandcorrectlyrepresentedtheagesofthosesiblingswouldreceivefullcredit.
4ManyoftheexamplesusedtoillustratetheColdStarttaskaredrawnfromTheSimpsonstelevisionshow.ReaderslackingadetailedworkingknowledgeofgenealogicalrelationshipsintheBouvier/Simpsonfamilyneednotagonizeoverwhattheyhavebeendoingwiththeirlivesforthepastquartercentury,butmaysimplyvisithttp://simpsons.wikia.com/wiki/Simpson_Family.
5
Input
Schema
TheKBschemaforColdStart2017consistsof:
• Entities:Entitiesandentitymentionsforfiveentitytypes(person,organization,geopoliticalentity,facility,andlocation)asdefinedinthetrilingualEDLtaskofthe2017EDLtrack.
• SFRelations:Entityattributes("slots")asdefinedintheSFtrack.Thesecompriseforty-onerelationtypesandtheirinverses.
• Events:Eventsandeventmentionsfor18eventsubtypes.ColdStartdefinesaneventasaRichEREcross-documenteventhopper,andaneventmentionasaRichEREeventtrigger.
• Event(Argument)Relations:Eventrolesandargumentsforthe18eventsubtypes,asdefinedintheEALtrack.Thesecomprise85eventargumentrelationtypeshavingeventasthepredicatesubject,and56inverseshavinganentityasthepredicatesubjectandaneventasthepredicateobject.
• SentimentRelations:Positiveandnegativesentimentfromasourceentitytowardatargetentity,asdefinedintheBeSttrack.
TheschemaforColdStart2017combinestheentityandmentiontypesfromTACKBP2016EntityDiscoveryandLinking,theSFrelationtypesfromTACKBP2016ColdStartKnowledgeBasePopulation,theeventrolesandargumenttypesfromTACKBP2016EventArgumentandLinking,theeventmentionsfromTACKBP2016EventNuggetDetection,andsentimentrelationsfromTACKBP2016BeliefandSentiment(BeSt).Annotation/assessmentguidelinesareavailableontheTACwebsite(http://www.nist.gov/tac/2017/KBP/ColdStart/guidelines.html),andaremorefullydocumentedinthedatapackagesthatcanberequestedfromtheLDCuponcompletionofTACKBPtrackregistration.
Forrelationswhoseobjects(“slotfills”)areentities(suchasper:siblingsorper:likes)orobjects(suchasper:confict.attack_attacker),ColdStartKBswillberequiredtolinkthatslottothenodeinthesubmittedKBrepresentingthecorrectentity5orevent.Slotswhosefillsarestrings(suchasper:titleororg:website)mustbelinkedtoaspeciallycreatedstringnodetorepresenttheobjectforthatrelation.
ColdStartentitiesandentitymentionsaredefinedbyDEFTRichERE.FullannotationguidelinesforDEFTRichEREentitiesareincludedintheDEFTRichEREannotationpackages,availablefromtheLDC,butahigh-levelsummaryofthefiveentitytypesandtheirmentionsareavailableinRichEREAnnotationGuidelinesOverview.ForColdStart,allnamedandnominalmentionsmustbeextracted,andtheentitiesmustbespecificindividualentities(asdescribedinAnnotationGuidelinesforIndividualityofSpecificEntities).Pronominalentitymentionsmaybeextracted(foruseinthesentimentcomponentevaluationinBeSt),butarenotrequiredforthecompositeevaluationoftheColdStartKB.AColdStartnamedentitymentionisthesameasanamedentitymentioninRichERE;i.e.,aColdStartnamedentitymentionisamentionthatuniquelyreferstoanentitybyitspropername,acronym,nickname,alias,abbreviation,orotheralternatename,andincludespostauthornamesfoundinthemetadataofdiscussionforumdocuments.Theextentof5Becausefacilityandlocationentitiesarenotincludedintheslotdefinitions,onlyperson,organization,andgeopoliticalentitynodesmustbelinkedtotheSFslots.
6
thenamedentitymentionistheentirestringrepresentingthename,excludingtheprecedingdefinitearticleandanyotherpre-posedorpost-posedmodifiers.AColdStartnominalentitymentionistheheadofthenominalentitymentioninRichERE;i.e.,aColdStartnominalentitymentionisamentionnotincludingtheentity'spropername,referringtoitbyacommonnounphrase(butforColdStart,thenominalmentionisonlytheheadnounofthenominalphrase).Entitymentionsareallowedtonestoroverlap;forexample,thestring“PhiladelphiaEagles”mightbeamentionofanORG(thefootballteam),whilethefirstwordmightsimultaneouslybeamentionofaGPE(thecityofPhiladelphia).
ColdStartdefinesaneventasaRichEREcross-documenteventhopper,andaneventmentionasaRichEREeventtrigger(whichisalsothesameasaneventnuggetintheTACKBPEventNuggettrack).Thecriteriafordeterminingwhethertwoormoreeventmentionsbelonginthesamehopperareessentiallythesameregardlessofwhetherthementionsareinthesamedocumentordifferentdocuments.TheColdStartinventoryofeventsandeventroles/argumentsisasubsetoftheeventtypesinRichERE,andisdescribedintheEAL2016Taskdescription.
TheColdStartinventoryofSFslotsisdescribedthoroughlyinTACKBP2015SlotDescriptionsandTACKBP2015AssessmentGuidelinesavailableontheTACWebsite.Forty-oneslotsandtheirinversesareusedfortheevaluation.Twenty-sixofthesehavefillsthatarethemselvesentities,asshowninTable1oftheAppendix.Theremainingfifteenslotshavestringfills,asshowninTable2oftheAppendix.EachSFpredicatehavinganentityasobjectwillhaveaninverse.6
ColdStartsentimentrelationsarealsogiveninatableintheAppendix,andincludeaninverseforeachsentimentpredicate.
ColdStarteventrelationsaregiveninthelasttwotablesintheAppendix.Thefirstshowseventpredicateshavinganeventassubject,andanentityorstringasobject.Thesecondtableshowstheinverseeventpredicates,whichhaveanentity(butnotastring)assubject,andaneventasobject.
Allinverserelationsmustbeexplicitlyidentifiedinthesubmittedknowledgebase.Thatis,iftheKBassertsthatrelationRholdsbetweenentitiesAandB,thenitmustalsoassertthatrelationR-1holdsbetweenBandA.Asaconvenience,theColdStartKBvalidationscriptcanbeusedtointroducemissinginversesintoaKB,exceptthatthevalidatorwillnotinferanyinverserelationsforeventpredicateshavingeventassubject;alleventrelationshavinganeventassubjectmustbeexplicitlyincludedintheKB.
DocumentCollection
TheColdStart2017evaluationdocumentcollectionwillbetheTACKBP2017EvaluationSourceCorpus,whichcomprisesapproximately90,000documents,roughlyequallydistributedbetweenEnglish,Spanish,andChinese,andbalancedbetweennewswire(NW)andmulti-postdiscussionforum(MPDF)documents.Thesedocumentswillbenew(previouslyunreleased)documentsthatwillbedistributedbyNISTviaWebdownloadatthebeginningoftheColdStartevaluationwindow.Therewillbeexactlyonefileperdocument,andallfileswillbeparsableasXML.Eachfilewillbegin
6SomeSFslots,suchasper:siblings,aresymmetric.Others,suchasper:parents,haveinversesthatwerealreadyinthe2014EnglishSlotFillingtrack(inthiscase,per:children).TheremainingSFslots(e.g.,org:founded_by)hadnocorrespondingslotinthe2014EnglishSlotFillingtrack;ColdStartspecifiesnewslotnamesfortheseinverses.Allsuchslotsarelist-valued.
7
withtheopeningtagoftheelement(forMPDF);7notethatcanbespelledwitheitheruppercaseorlowercaseletters,dependingonthegenre,andmayoptionallyincludeadditionalattributes(suchas"type"forsomenewswiredata).
Newswiredatawillusethefollowingmarkupframework:
...
...
...
...
wheretheHEADLINEandDATELINEtagsareoptional(notalwayspresent),andtheTEXTcontentmayormaynotinclude"
...
"tags(dependingonwhetherornotthe"doc_type_label"is"story").
Multi-PostDiscussionForumfiles(MPDFs)arederivedfromDiscussionForumthreads.Theyconsistofacontinuousrunofpostsfromathreadbuttheyareonlyapproximately800wordsinlength(excludingmetadataandtextwithinelements).Whentakenfromashortthread,aMPDFmaycomprisetheentirethread.However,whentakenfromlongerthreads,aMPDFisatruncatedversionofitssource,thoughitwillalwaysstartwiththepreliminarypost.TheMPDFfileswillusethefollowingmarkupframework,inwhichtheremayalsobearbitrarilydeepnestingofquoteelements,andotherelementsmaybepresent(e.g."..."anchortags):
...
... 7IncontrasttosomeoftheKBPsourcecorporapriorto2016,theTACKBP2017SourceCorpuswillnotcontainanyfilesthatbeginwithxmldeclarationssuchas.ThisistoensurethatoffsetsalignacrossthevariousKBP2017tracksthatareusingthissameevaluationsourcecorpus,regardlessofwhetheroffsetsarecountedfromthebeginningofthefile,orthebeginningofthetag.
8
...
...
...
Allprovenance/justificationsforColdStartKB/SF2017tasksmustbedrawnfromthedocumentsintheTACKBP2017EvaluationSourceCorpus.EachdocumentisrepresentedasaUTF-8characterarrayandbeginswiththetag,wherethe“
9
AllCSSEevaluationqueriesstartwithanentrypointintotheknowledgebasebeingevaluated.Theentrypointisdefinedbyanamedentitymention(name,docid,beginoffset,andendoffset),andisfollowedbytheentitytypeandeitheroneortwoslotstobeextractedfortheentity.ThequerymayrequestanyoftheSF,Sentiment,orEventrelationsthathavethequeryentityasthepredicatesubject.
Evaluationqueriescouldtakeoneoftwoforms:single-hopormultiple-hop.Forexample,hereisasamplesingle-hopCSSEevaluationquerythatasks“WhatistheageoftheJuneMcCarthymentionedatoffsets16931-16943inDocument42?”:
June McCarthy 42 16931 16943 PER per:age per:age
Thissingle-hopquerylooksverymuchlikeaqueryfromthe2014EnglishSlotFillingtask,exceptthateachqueryinColdStartasksforaspecificslot,ratherthanallslotsforwhichthereisinformationinthedocumentcollection.8
Amorecomplex“two-hop”querymightask,“WhataretheagesofthechildrenoftheJuneMcCarthymentionedatoffsets16931-16943inDocument42”:
June McCarthy 42 16931 16943 PER per:children per:children per:age
TheabovequeriesarehomogeneousSFqueriesinthesensethateachqueryasksforonlySFrelations.Anexampleofamixedquery,involvingbothasentimentpredicate(inslot0)andaneventpredicate(inslot1)isbelow(“WhataretheattackeventsinwhichthetargetisapersonwhomJuneMcCarthydislikes?”):
June McCarthy 42 16931 16943 PER 8ParticipantsintheSlotFillingvariantshouldtreatallotherslotsasiftheyappearinthefieldofaSlotFillingqueryfrom2013orearlier.
10
per:dislikes per:dislike per:conflict.attack_target
Ingeneral,two-hopquerieswillstartfromanentrypoint(selectingthecorrespondingKBentityofaCSKBsubmission),followasingleentity-valuedrelation,thenaskforasingleslotvalue.9Suchquerieswillverifythattheknowledgebaseiswell-formedinawaythatgoesbeyondbasicentitylinkingandslotfilling,withoutallowingcombinationsoferrorstodrivescorestozero.
Becausetwo-hopqueriesdonotlooklikeanyslotfillingqueriesfromKBP2009-2014,participantsintheColdStartSlotFillingvariantmustprocesstheCSSFqueriesintwo“rounds”usingtheCS-GenerateCSQueries.plscriptfromNIST,whichaddstheentrytotheNIST-distributedCSSFqueries.ParticipantsintheSlotFillingvariantmusttreatastheslottobefilled.Duringthefirstround,willbeidenticalto.TheCS-GenerateCSQueries.plscriptwillthenconvertafirstroundoutputfiletoasecondroundqueryfile.Secondroundqueriesgeneratedbythisscriptwillbearentriesequivalentto.ThoughsomeoftheCSSFquerieswilldifferonlyinhavingdifferentmentions(possiblyforthesameentity)astheirentrypoints,participatingCSSFsystemsareprohibitedfromusinginformationaboutonequerytoinformtheprocessingofanotherquery.
FortheKnowledgeBasevariant,thefollowingrulesareappliedtomapfromaCSSEevaluationquerytoaknowledgebaseentry:First,formacandidatesetofallKBnodementionsthathaveatleastonecharacterincommonwiththeevaluationquerymentionandthathavethesametype.Ifthissetisempty,thesubmissiondoesnotcontainanyanswersfortheevaluationquery.Otherwise,foreachmentionKinthecandidateset,calculate:
• COMMON,thenumberofcharactersinKthatarealsointhequerymentionQ.• K_ONLY,thenumberofcharactersinKthatarenotinQ.
Executeeachthefollowingeliminationsuntilthecandidatesetissizeone,andselectthatcandidateastheKBnodethatmatchesthequery:
• EliminateanycandidatethatdoesnothavethemaximalvalueofCOMMON• EliminateanycandidatethatdoesnothavetheminimalvalueofK_ONLY• Eliminateallbutthecandidatethatappearsfirstinthesubmissionfile
TheproperspecificationofentitymentionsinaKBisthereforeimportantforscoringwell;CSKBparticipantsshouldthereforetakecaretoensurethateverynamedentitymentionintheevaluationcollectionservesasamentionforanodeintheKB.
TheNISTevaluationofaKBwillproceedbyfindingallentriesintheKBthatfulfillanevaluationquery.Forexample,iftheevaluationquery‘schoolsattendedbythesiblingsofBartSimpson’findstwosiblingsforthenodespecifiedbytheentrypoint,andtheKBindicatesthatthosesiblingsattendedtwoandoneschoolsrespectively,thenthreeresultswouldbeassessedbyNIST.TheseresultswillbeconvertedtothesameformastheoutputfortheSlotFillingvariant.ResultswillbepooledacrossallCSKBandCSSFsubmissions,andassessorswilljudgethevalidityofeachresult.Finally,ascoringscriptwillreportavarietyofstatisticsforeachsubmittedrun.
9Inprinciple,multiple-hopqueriescouldincludemorethantworelations,butwecurrentlylimitourselvestotwo.
11
Increatingevaluationqueries,LDCwillexercisearangeofSF,sentiment,andeventpredicatesandstrivetobalanceevendistributionacrosspredicatetypeswithproductivityofthoseslots.Thismeansthatthequeriesinthecompositeevaluationwillnotnecessarilyfollowthedistributionofmention-leveloccurrencesoffacts,soevenifthereare10timesasmanynegativesentimentsaspositivesentiments,thenumberofqueriesaskingforpositivesentimentvsnegativesentimentwillbeaboutthesame.Systemsthathavebeenoptimizedforthecomponentevaluationsbyassumingaparticulardistributionofmention-levelphenomenainthecomponentevaluationdocuments,mayneedtorecalibratetotakeintoaccountlessfrequentphenomena(eventtypes,cognitivestates,etc.).
Singlehopquerieswillinmanycasesaskformultipleslotsforagivenentityregardlessofwhetherfillersforthoseslotsareattestedinthedocumentcollection.Multiplehopquerieswillfavorentitiesandslotsequencesthatareattestedinthedocumentcollection(althoughheretoo,availabilityofanswersisnotguaranteedatanyhoplevel).
Becausecoreferenceofstringsandeventsisstilldifficult(forbothhumansandautomaticsystems),ColdStart2017willmitigateerrorsandinter-annotatordisagreementabouthowtocoreferenceeventsandstrings(especiallycross-documentcoreference)byavoidingqueriesthatinvolveaneventorstringasthepredicatesubject.ColdStart2017querieswillalwayshaveanentityasthepredicatesubjectateachhoplevel.Thismeansthatalleventquerieswillbeoftheform:FindalleventsthathaveentityXinroleY(e.g.,“Findallconflict.attackeventsthathavetheentity“HomerSimpson”intherole“attacker”,butnot“Findalltargetsofconflict.attackeventswhereHomeristheattacker”).
Single-hopquerieswillrequestsomeeventorstringobjectinslot0;two-hopquerieswillrequestanentityobjectinslot0,andanentity,event,orstringobjectinslot1.Thetargeteddistributionofpredicatesacrossqueriesthatareassessedinthefinalevaluationisgiveninthefollowingtable.
ApproximatedistributionofpredicatetypesacrossColdStartevaluationqueries
Single-hopqueries Two-hopqueries
Slot0 event SF(stringobject) SF SF SF sentiment sentiment sentiment
Slot1 -- -- SF sentiment event SF sentiment event
3/12 1/12 3/12 1/12 1/12 1/12 1/12 1/12
ColdStartKBTaskOutputTheColdStartknowledgebaseisrepresentedasadirectedlabeledmultigraph:entities,events,andotherpredicatearguments(i.e.,“string”arguments,thatareneitherentitiesnorevents)arerepresentedasnodesintheknowledgegraph,whilebinaryrelationsbetweentheentities/events/stringsarerepresentedasedgeslabeledwithSF,sentiment,andeventpredicatenames.Generally,thesubjectandtheobjectofthepredicatearebothnodesintheKB(exceptthatthetype,link,and*mentionpredicatestakeaquotedstringasobject).TheKBisgroundedtothedocumentcollectionviavarious*mentionpredicatesthatconnecteachentity,event,orstringnodetoitsmentionsinthedocumentcollection.TheColdStartKBisconnectedtoexternalKBsviaalinkpredicatethatindicatescoreferencebetweenanentityintheColdStartKBandanentitynodeintheexternalKB.
12
CSKBsystemsmustproduceaknowledgebaseasoutput.ThefirstlineoftheKBoutputfilemustcontainauniquerunID.TheremainderoftheKBoutputfileisasetofassertions,oraugmentedtriples(subject,predicate,object).Assertionswillappear,one-per-line,intab-separatedformat.TheKBoutputfilewillbeautomaticallyconvertedtoRDFstatementsduringevaluation.AllKBoutputmustbeencodedinUTF-8.
EachtripleappearsintheKBoutputfileinsubject-predicate-objectorder.Forexample,toindicatethatEntity4hasEntity7asasibling,thetriplemightbe:
:Entity4 per:siblings :Entity7
IfEntity4hassiblingsinadditiontoEntity7,theserelationsshouldbeenteredasseparatetriples.
EachtripleintheCSKBsubmissionwillincludeasetofaugmentations(againusingtabsasseparators).Exceptforthetypepredicate(whichdoesnotrequireexplicitsupportfromadocument)thefirstaugmentationwilldescribetheprovenance(ajustification)ofthetriple,andthesecondaugmentationwillprovidetheconfidenceforthetripleandjustification.Ifthereismorethanonejustificationforatriple,eachjustificationappearsinaseparateassertion(line),alongwithitsconfidence.
Atleastoneassertionforeachuniquesubject-predicate-objecttriplewillbeevaluated.Ifmorethanoneassertionofagiventripleappearsintheoutput(witheachtriplehavingdifferentprovenance),LDCwillassesstheassertionwiththehighestconfidencevalue(seebelow),andwillassessadditionalassertionsifresourcesallow.Ifmorethanonesuchassertionsharesthesameconfidencevalue,theassertionthatappearsearlierintheoutputwillbeconsideredtohavehigherconfidence.
Nodes
TheKBcontainsthreedifferentkindsofnodes:Entity,Event,andString.Eachnodespecificationbeginswithoneof“:Entity”,“:Event”,or“:String”,followedbyasequenceofletters,digits,andunderscores.Examplesoflegalentity,event,andstringnodespecificationsinclude:Entity42, :Event_056,and:String74_R29,respectively.Nomeaningisascribedtothissequencebytheevaluationsoftware;itisusedonlyasauniqueidentifier.Anysubsequentuseofthesamecolon-precededsequencewillbetakenasareferencetothesameentity,event,orstringnode.
Eachspecificindividualentityoreventthatappearsinthedocumentcollectionmustberepresentedbyexactlyoneentityoreventnode.Twoseparateentity(orevent)nodesintheKBwillbeinterpretedasrepresentingtwodifferententities(orevents).
Thestringnodeisacatch-allstructuretorepresentSFandEventpredicateargumentsthatarenotspecificindividualentities.ThestringnodeallowstheKBtorepresentmultiplejustificationsforthesamesubject-predicate-objecttriplewhentheobjectisstring-valued(e.g.,per:cause_of_death),evenwhentheliteralstringsaredifferent(e.g.,“cardiacarrest”vs.“heartattack”);ifagivensubjectandpredicatehastwoseparatestringnodesasobjects,theywillbeinterpretedasrepresentingtwodifferentslotfillersforthatsubjectandpredicate.However,inTAC2017thereisnorequirementthatthesamestringnodebeusedfortripleshavingdifferentsubjectsorpredicates(i.e.,for"JohnandMarydiedofheartdisease",theKBmaycontaintwodifferentstringnodesfor"heartdisease",whereonenodeisthecauseofdeathofJohn,andtheothernodeisthecauseofdeathofMary).10
10TheKBmayelecttorepresenteachreal-worldconceptorvaluebyexactlyonenode(e.g.,fordatesandnumericvalues),butsuchaglobalgroundingofnodestotherealworldisnotarequirementinTAC2017(beyondtheentitiesandeventsthataredefinedexplicitlyinthe2017TACKBPontology).Aricherontologyofconceptsandvaluesisleftforfuturework.
13
Forthe2017evaluation,thestringobjectonlyneedstoincludeonementionforeachdocumentprovidedasjustificationforthatsubject-predicate-objecttriple.
Inparticular,eacheventargumentthatisnotaspecificindividualentity(e.g.,unnamedaggregateslike“3people”)mustberepresentedintheColdStartKBasastringnode.WhentheeventargumentisnotavalidColdStartentity(i.e.,aspecificindividualPER,ORG,GPE,LOC,orFAC),theKBmayrepresenteachargumentstringasaseparatestringnode,evenifit’sclearfromcontextthatthestringsarecoreferential;stringvaluedeventargumentsareignoredinthecompositeKBevaluation(ColdStart2017eventquerieswillalwayshaveanentitynodeassubject,andaneventnodeasobject),andareusedonlytoproducetheEALoutputfilesforthecomponentEALevaluation(whichdoesnotrequireexplicitcoreferenceofeventargumentstrings).
Predicates
ThelegalpredicatesaretheSlotFillingpredicates,sentimentpredicates,andeventpredicatesshownintheAppendix,plustype,link,mention,nominal_mention,pronominal_mention,normalized_mentionandcanonical_mention.
SFpredicatesfoundinTable1musthaveentityspecificationsinboththesubjectandobjectpositions;predicatesfoundinTable2mustspecifyanentitynodeinthesubjectposition,andastringnodeintheobjectposition;thestringnodeintheobjectpositionwillexactlycorrespondwiththeslotfillforthatrelationintheSlotFillingtask.
type
Eachentity,event,andstringnodewillbethesubjectofexactlyonetypetriple.TheobjectofthattriplewillbeoneoftheallowabletypeslistedinTable1below.Itisuptosubmittingsystemstocorrectlyidentifyandreportthetypeofeachentityandevent;allstringnodesmusthavetypeSTRING.
Table1Allowablevaluesfortypepredicate
Node Allowabletype
String STRING
Entity PER
ORG
GPE
LOC
FAC
Event CONFLICT.ATTACK
CONFLICT.DEMONSTRATE
CONTACT.BROADCAST
14
CONTACT.CONTACT
CONTACT.CORRESPONDENCE
CONTACT.MEET
JUSTICE.ARREST-JAIL
LIFE.DIE
LIFE.INJURE
MANUFACTURE.ARTIFACT
MOVEMENT.TRANSPORT-ARTIFACT
MOVEMENT.TRANSPORT-PERSON
PERSONNEL.ELECT
PERSONNEL.END-POSITION
PERSONNEL.START-POSITION
TRANSACTION.TRANSACTION
TRANSACTION.TRANSFER-MONEY
TRANSACTION.TRANSFER-OWNERSHIP
*mentionpredicates
Eachentity,event,andstringnodewillbethesubjectofoneormorepredicatesfrom{mention, nominal_mention, pronominal_mention, canonical_mention, normalized_mention};theterm“*mention”isusedtorefertothesepredicates.Togetherwiththeprovenanceinformation(seebelow),these*mentiontriplesindicatehowtheknowledgebaseistiedtothedocumentcollection.Theobjectofa*mentiontripleisthedouble-quotedmentionstring;documentIDandoffsetappearunderprovenanceinformation(seebelow).
mention and nominal_mention and pronominal_mention
Eachentitywillbethesubjectofoneormoremention,nominal_mentionorpronominal_mentiontriples.Thedefinitionofwhatconstitutesanamed,nominal,orpronominalentitymentionforColdStartisdescribedintheColdStartschemaabove.Eachnamedentitymentioninthecollectionmustbesubmittedastheobjectofamentiontriple,whileeachnominalentitymentioninthecollectionmustbesubmittedastheobjectofanominal_mentiontriple.Forexample,ifanentityismentionedbynamefivetimesinadocument,fivementiontriplesshouldbegenerated.Thepronominal_mentionsareusedonlyforthecomponentBeStevaluation,andarenotreferencedor
15
evaluatedinthecompositeKBevaluationoranyoftheothercomponentevaluationsbesidesBeSt;therefore,theKBonlyneedstoincludepronominalmentionsthatserveasprovenanceforsentimentassertions.Anexampleisshownbelowtodemonstratetheusageoftheassertion:
:Entity_0007 type PER
:Entity_0007 canonical_mention "Dzhokhar Tsarnaev" NYT_ENG_20131113.0264:434-450 1.0
:Entity_0007 mention "Dzhokhar Tsarnaev" NYT_ENG_20131113.0264:434-450 1.0
:Entity_0007 pronominal_mention "he" NYT_ENG_20131113.0264:546-547 1.0
Eacheventwillbethesubjectofoneormorementiontriples.ThedefinitionofwhatconstitutesaneventmentionisdescribedintheColdStartschemaabove.Eachmentionoftheevent(i.e.,eacheventnuggetorEREeventtriggerfortheevent)mustbesubmittedastheobjectofamentiontriple.EventmentionsneedtobeexhaustiveinordertosupportevaluationofthecomponentEventNuggetDetectionandCoreferencetask.
Eachstringnodewillbethesubjectofoneormoremorementiontriples.Thestringnodeonlyneedstoincludeonementionforeachdocumentprovidedasjustificationforthatsubject-predicate-objecttriple.
canonical_mention
Foreachdocumentthatmentionsanentityorevent,oneofthementions(ornominal_mentionsifit’sanentity)mustbeidentifiedasthecanonicalmentionforthatentity/eventinthatdocument;itisthestringthatwillbeseenbytheassessorifthatentity/eventappearsasaslotfill,supportedbythatdocument(inSlotFillingtaskterms,itisthecontentofColumn5ofaCSSF2017submission,anditsprovenancewillserveasColumn7oftheCSSFsubmission).13Thisimpliesthatadocumentattestingtoarelationmustcontainmentionsornominal_mentionsofboththesubjectandtheobjectoftherelation.Canonicalmentionsareexpressedusingacanonical_mentiontriple.Theargumentsforcanonical_mention arethesameasformention.Notethatthereisnorequirementthatsubmissionsselectasingle,globalcanonicalmentionforanentity.Whilesuchamentionmightbeuseful,herewerequirethatacanonicalmentionbeprovidedwithineachdocumentfortheassessortouseduringassessment.
Eachcanonical_mentionisalsoamention (or nominal_mention or pronominal_mentionifthenodeisanentity).Asaconvenience,ifasubmittedKBdoesnotcontainamention(ornominal_mention/pronominal_mention)tripleforeachcanonical_mentiontriple,themissingrelationswillbeinferred(perhapsincorrectly)asmentions(albeitwithawarning).ThisshortcutisprovidedtomakesubmittedKBseasiertoview,anddoesnotrelievesubmittersfromtherequirementtoprovideeachoftherequiredmentions,nominal_mentions,andcanonical_mentions.
13IntheSlotFillingtaskofKBP2009-2014(andintheSlotFillingvariantofColdStart),allslotfillsarestrings.Assessorsverifythevalidityofaslotfillbylookingforthatstringinthespecifieddocument,usingtheprovenanceinformationprovidedinthesystemresponse.InasubmittedKB,slotsthatarefilledwithentitiesoreventsholdnotstrings,butpointerstotheKBstructurefortheappropriateentity/event.Thus,acanonicalmentionmustbeidentifiedbytheColdStartKBforeachentityineachdocument,sothattheassessorcanbepresentedwithastringthatrepresentstheentityduringassessment.
16
normalized_mention
Inordertoallownormalizeddates(andothernormalizedstringsinfuture)intheKB,astringnodeforanormalizedstringvaluemustbethesubjectofanormalized_mentionpredicate.Anexampleoftheusageofanormalized_mentionassertionisshownbelow:
:String_0001 type STRING
:String_0001 mention "April 15" NYT_ENG_20131113.0264:624-631 1.0
:String_0001 normalized_mention "2013-04-15" NYT_ENG_20131113.0264:624-631 1.0
Thestringprovidedastheobjectofnormalized_mentionwouldnotbeverifiedagainsttextinthesourcedocument;however,itisarequirementthatforagivenstringnodetheprovenanceofthenormalized_mentionassertionshouldbethesameastheprovenanceofanother(non-normalized)mentionofthatstring.
link
Eachentitymaybethesubjectofuptoonelinkpredicate.Theobjectofthepredicateisaquotedstringoftheform“ExternalKBID:ExternalNodeID”andindicatesthattheColdStartentityisthesameastheentitywithID“ExternalNodeID”inanexternalreferenceKB.ForTAC2017,theexternalreferenceKBisthesameasthatusedin2015-2017fortheTACTrilingualEDLtrack,namelyLDC2015E42(TACKBPKnowledgeBaseII–BaseKB).ThelinkpredicateisignoredinthecompositeKBevaluationandisusedonlyfortheEDLcomponentevaluation.
Thefollowingexampleshowshowtouselink(assumingtheexternalKBisthereferenceKBinLDC2015E42:TACKBPKnowledgeBaseII-BaseKB):
DEMO
:Entity_0001 type GPE
:Entity_0001 canonical_mention "Boston" NYT_ENG_20131113.0264:402-407 1.0
:Entity_0001 mention "Boston" NYT_ENG_20131113.0264:402-407 1.0
:Entity_0001 link "LDC2015E42:m.050v43"
ThiswillproducethefollowingEDLoutput:
DEMO :Entity_0001_M00001 Boston NYT_ENG_20131113.0264:402-407 m.050v43 GPE NAM 1.0
SF,sentiment,andeventpredicates
TheKBmustincludealltriplesinvolvingSFpredicates,sentimentpredicates,andeventpredicatesintheAppendix.
EventRealis
TheKBmustspecifyrealisfor
• eventpredicates,and• *mentionpredicatesthathaveaneventnodeassubject.
Realismaytakeoneofthefollowingvalues:
• actual
17
• generic• other
InordertosupporttheEventNuggetandEventArgumentandLinkingcomponentevaluationsoftheKB,theKBmustspecifytherealisofeventmentionsandeventargumentassertions,andincludeeventsandargumentassertionsofallthreerealisvalues.TherealisofaneventmentionintheColdStartKBfollowsthedefinitionofrealisinRichERE,whiletherealisofaneventargumentassertionfollowsthedefinitionintheEventArgumentandLinkingtask(EAL).
FromtheperspectiveofaColdStartKBuser,bothcompletedandplannedeventsareofinterest,butnotgenericevents;therefore,whenaColdStartqueryrequestinganeventisappliedtotheKB,itwillconsideronlyeventnodesthathavean“actual”or“other”mention,wherethequeryentityisan“actual”or“other”argumentfortheevent.Thismeans,forexample,thatthequery"FindallattackeventsthathaveHomerasanattacker"means"FindallACTUALorOTHEReventsthathaveHomerasanACTUALorOTHERattacker".GENERICevents(thathavebeenmis-classifiedasACTUALorOTHERintheKB)willbeassessedasWrong.
Eventsthathave“generic”realisintheKBareignoredinthecompositequery-basedevaluation,andareusedonlyforthecomponenteventnuggetandeventargumentevaluations.
Therealisvalueshouldbeappendedattheendofthepredicatename,using"."toseparatethetwo.Forexample,
:Event_0001 type CONFLICT.ATTACK
:Event_0001 mention.actual "bombing" NYT_ENG_20131113.0264:418-424 1.0
:Event_0001 canonical_mention.actual "bombing" NYT_ENG_20131113.0264:418-424 1.0
:Event_0001 conflict.attack:attacker.actual :Entity_0007 NYT_ENG_20131113.0264:492-681;NYT_ENG_20131113.0264:546-547;NIL 1.0
Provenance
Eachassertion(exceptfortypeassertions)mustcontainasinglejustification(provenance)immediatelyafterthesubject-predicate-objecttriple.Provenanceisasetofjustificationspans;eachspanmaycompriseatmost200UTF-8characters.EachjustificationspanwillincludeadocumentID,followedbyacolon,followedbytwodash-separatedoffsets(beginandendoffsets).Theoffsetsthatshowtheprovenanceofanextractedrelationareusedtonarrowtheassessor’sfocuswithinthedocumentswhenassessingthecorrectnessofthatrelation.
Provenancespansforasinglejustificationmustcomefromasingledocument.Thisisanewrestrictionin2017,toallowjustificationstobecountablebasedonthejustificationdocuments.
Provenancespanscanbepartitionedintofourdifferentgroups:
• FILLER_STRING(musthaveexactly1span)• PREDICATE_JUSTIFICATION(mayhave1-3spans;multiplespansareseparatedbya
comma)• BASE_FILLER(musthaveexactly1span)• ADDITIONAL_JUSTIFICATION(mayhaveanynumberofspans;multiplespansare
separatedbyacomma)
Provenancefortheassertionconsistsofsomesubsetofthefourspangroupsabove,dependingonthepredicateandobject;multiplegroupsareseparatedbysemicolon.
18
a) Ifthepredicateistype:- Noprovenanceshouldbeprovided
b) Otherwise,ifthepredicateisanyofthe*mentionpredicates:- ProvenanceconsistsofonlyPREDICATE_JUSTIFICATION,containingexactlyone
span,representingtheexactlocationofthementioninthedocumentcollectionc) Otherwise,ifthepredicateisasentimentpredicate:
- ProvenanceconsistsofonlyPREDICATE_JUSTIFICATION,containingexactlyonespan
- Theprovenancemustbeamentionoftheentitythatisthetargetofthesentiment,andmustbethementionclosesttowherethesentimentisexpressed
- N.B.:Thetargetofthesentimentcouldbeeitherthesubjectorobjectofthepredicate,dependingonthepredicate:
- forlikesanddislikespredicates,thetargetofthesentimentistheobject- foris_liked_byandis_disliked_bypredicates,thetargetofthesentimentisthe
subjectd) Otherwise,ifthepredicateisanSFpredicatewithanon-stringobject:
- ProvenanceconsistsofonlyPREDICATE_JUSTIFICATION,containing1-3spanse) Otherwise,ifthepredicateisanSFpredicatewithastringobject:
- ProvenanceconsistsofFILLER_STRING;PREDICATE_JUSTIFICATION- FILLER_STRINGmustbeoneofthementionsofthestringobject- PREDICATE_JUSTIFICATIONcontains1-3spans
f) Otherwise,ifthepredicateisaneventpredicatewithanon-stringobject- Provenanceconsistsof
PREDICATE_JUSTIFICATION;BASE_FILLER;ADDITIONAL_JUSTIFICATION- PREDICATE_JUSTIFICATIONcontains1-3spans- ADDITIONAL_JUSTIFICATIONmaybe"NIL"oranynumberofspans
g) Otherwise,ifthepredicateisaneventpredicatewithastringobject- Provenanceconsistsof
FILLER_STRING;PREDICATE_JUSTIFICATION;BASE_FILLER;ADDITIONAL_JUSTIFICATION
- FILLER_STRINGmustbeoneofthementionsofthestringobject- PREDICATE_JUSTIFICATIONcontains1-3spans- ADDITIONAL_JUSTIFICATIONmaybe"NIL"oranynumberofspans
Forpredicateswithastringobject,thefirstjustificationspan(FILLER_STRING)mustrepresentthedocumentIDandoffsetsofthestringfill.(SlotFillingvariantparticipantsarealreadyprovidingthisinformationinColumn7oftheirsubmissions.)Thisisthetextthatwillbeshowntoassessorsinsteadofacanonical_mentionforthestringnode.
ConfidenceMeasure
Topromoteresearchintoprobabilisticknowledgebasesandconfidenceestimation,eachassertionintheKB(orslotfillintheCSSFsubmission)musthaveanassociatedconfidencescore.ConfidencescoreswillbeusedtocomputeavariantofMeanAveragePrecision(MAP)inthecompositeKBevaluation.Confidencescoreswillbeusedtoinduceatotalorderovertherelationsbeingevaluated(tiesarebrokenwhentwoscoresareequalbyassumingthattheassertionappearingearlierinthesubmissionhasahigherscore).Anysubmittedconfidencescoremustbeapositiverealnumberbetween0.0(exclusive,representingthelowestconfidence)and1.0(inclusive,representingthehighestconfidence),andmustincludeadecimalpoint(nocommas,please)toclearlydistinguishitfromadocumentoffset.Confidencescores,ifpresent,willappearattheendofeachoutputline,separatedfromtheprovenanceinformationwithatab.Ifnoconfidencescoreisprovidedforan
19
assertion,theconfidencewillbeinferredtobe1.0forthepurposesofevaluation.Confidencescoresmaynotbeusedtoqualifytwoincompatiblefillsforasingleslot;submittersystemsmustdecideamongstsuchpossibilitiesandsubmitonlyone.Forexample,ifthesystembelievesthatBart’sonlysiblingisLisawithconfidence0.7andMilhousewithconfidence0.3,itshouldsubmitonlyoneofthesepossibilities.Ifbotharesubmitted,itwillbeinterpretedasBarthavingtwosiblings.
Comments
Outputfilesmaycontaincomments,whichbeginatanyoccurrenceofapoundsign(#)andcontinuethrough(butdonotinclude)theendoftheline.Commentsandblanklineswillbeignored.ThefirstlineofaKBvariantoutputfilemustcontaintheuniquerunID(i.e.,itmaynotbeblank).Submittersmayliketoaddacommenttothislinegivingfurtherdetailsabouttherun.
SlotFillingTaskOutputOutputfortheSlotFillingvariantwillbeintheformofatab-separatedfile.Thecolumnsofthesubmittedfileareasfollows:
Column1 QueryID.Forthefirstround,thisistakendirectlyfromtheXMLtag.Forthesecondround,thisisdrawnfromthetagofthequerygeneratedfromthefirstroundoutput.
Column2 Thenameoftheslotbeingfilled.
Column3 AuniquerunIDforthesubmission.
Column4 Provenancefortherelationbetweenthequeryentityandslotfiller,consistingofupto3docid:startoffset-endoffsettriplesseparatedbycommas.Individualspansmaycompriseatmost200UTF-8characters.Unlikethe2014SlotFillingtask,thereisnorequirementtogenerateNILentrieswhennoinformationaboutthetargetentityisavailable.
Column5 Aslotfiller(possiblynormalized,e.g.,fordates).Thisisusedbothtopopulatetheentryofthenextroundquery,andbytheassessortojudgetheslotfill.ThestringshouldbeextractedfromthefillerprovenanceinColumn7,exceptthatanyembeddedtabsornewlinecharactersshouldbeconvertedtoaspacecharacteranddatesmustbenormalized(therefore,slotfillersshouldnotbetranslatedacrosslanguages).Ifanominalmentionisreturnedasaslotfiller,onlytheheadwordofthenominalphraseshouldbereturned(consistentwiththeEDLdefinitionofnominalmentions).Fordates,systemsmustnormalizedocumenttextstringstostandardizedmonth,day,and/oryearvalues,followingtheTIMEX2formatofyyyy-mm-dd(e.g.,documenttext“NewYear’sDay1985”wouldbenormalizedas“1985-01-01”);ifafulldatecannotbeinferredusingdocumenttextandmetadata,partialdatenormalizationsareallowedusing“X”forthemissinginformation.
Column6 Afillertype,selectedfrom{PER,ORG,GPE,STRING}.TheSTRINGfilleris
20
usedforstring-valuedslotsshowninTable2.
Column7 Provenancefortheslotfillerstring.Thisiseitherasinglespan(docid:startoffset-endoffset)fromthedocumentwherethecanonicalslotfillerstringwasextracted,or(inthecasewhentheslotfillerstringinColumn5hasbeennormalized)asetofuptotwocomma-separateddocid:startoffset-endoffsetspansforthebasestringsthatwereusedtogeneratethenormalizedslotfillerstring.ThedocumentsusedfortheslotfillerstringprovenancemustbeasubsetofthedocumentsprovidedinColumn4.Thiscolumnservestwopurposes.First,LDCwilljudgeCorrectvs.Inexactwithrespecttothedocument(s)providedintheslotfillerstringprovenance.Second,thiscolumnisusedtofillthe,andentriesinsecondroundqueries.Ifmorethanoneprovenancetripleisprovidedhere,thefirstonewillbeusedtofillthesecondroundquery.
Column8 Confidencescore.
TheprocessforconstructingaSlotFillingvariantsubmissionisasfollows:
• DownloadthefollowingfromtheNISTWebsite:o TheColdStartevaluationdocumentsCS-GenerateQueries.plscripto CS-PackageOutput.plscripto CS-ValidateSF.plscript
• [email protected]:o TheCSSFevaluationqueries
• ConfigureyoursystemtoproduceresultsonlyfromtheColdStartevaluationdocuments.• RuntheCS-GenerateQueries.plscriptontheevaluationqueriestoproducethefirstround
queriesyoursystemwillrunon.Notethattherawevaluationqueriesmightdifferfromtheformatgivenabove,soyoushouldnotassumethatyoucanusethemasinputtoyoursystemwithoutrunningthisscript.
• Runyoursystem,producingaslot-fillingsubmissionforthefirstroundqueries.• RuntheCS-ValidateSF.plscriptonyourfirstroundoutputtoverifythatitisformatted
correctly.• RuntheCS-GenerateQueries.plscriptontheevaluationqueriesandyourfirstround
outputtoproducethesecondroundqueries.• Runyoursystemonthesecondroundqueriestoproduceasecondoutputfile.• RuntheCS-PackageOutput.plscriptonthetwooutputfilestoproduceyoursubmission.• RuntheCS-ValidateSF.plscriptonyoursubmissiontoverifythatitisformattedcorrectly.• UploadthesubmissiontoNIST.
Slotfillingsystemsthatparticipatedinthe2014SlotFillingtaskwillneedtohandlethefollowingdifferencestosuccessfullyparticipateinthe2017CSSFtask:
• Onlytheslotspecifiedbytheentryistobefilled;allotherslotsshouldbeignored.TheentryisaddedtothequeriesreceivedfromNISTbyrunningtheCS-GenerateQueries.plscript.
• Participantswillneedtodooneroundofslotfilling,runtheCS-GenerateQueries.plscripttocreatethesecondroundqueries,thenrunslotfillingagainonthenewqueries.The
21
resultsofroundsoneandtwoaretobeconcatenatedbeforesubmissionusingtheCS-PackageOutput.plscript.
• CSSFrequiresthatparticipantsbeabletofillallslotsinbothdirections.Forexample,the2014SlotFillingtaskrequireddetectionofthe per:cities_of_residence slot.CSSFalsorequiressystemstobeabletodetecttheinverseofthatslot,gpe:residents_of_city.
• Eachslotfillermustbeassignedatype,selectedfrom{PER,ORG,GPE,STRING}.Thisfieldrepresentsanadditionaloutputcolumnnotfoundinthe2014SlotFillingorCSSFtasks.
• NILentries,indicatingthatnoinformationaboutaparticularslotisavailable,arenotrequiredinCSSF.
• Nominalmentionsofslotfillersmaybereturnifnonamedentitymentionisavailableinthedocumentcollection.(Returningnominalentitymentionsisnotrequired,butmayimprovesystemrecallifdonecorrectly.)
• ToconformwithrequirementsintheColdStartKBtask,provenanceforeachSFrelationislimitedtoonly3spans(insteadof4),andeachspanmayhaveupto200UTF-8characters(insteadof150).
HereareexamplelinesfromaSlotFillingsubmission:
Q4 org:city_of_headquarters myrun1 Doc42:3-8,Doc8:3-11 Baltimore GPE Doc8:3-11 1.0 Q5 per:siblings myrun1 Doc124:283-288,Doc885:173-179 Lisa PER Doc124:283-286 0.7
Q6 per:age myrun1 Doc124:180-181,Doc885:173-179 10 STRING Doc124:180-181 0.9
Evaluation
The submitted Cold Start KBs are evaluated by both a composite query-based evaluation, and a set of component evaluations. The composite KB evaluation applies a set of Cold Start evaluation queries to each KB and assesses the correctness of the events, sentiment sources and targets, and SF slot fillers found.
Becausethecompositeevaluationmayhidemanyfactorscontributingtotheperformanceoftheend-to-endKBsystem,eachsubmittedKBalsoundergoesasetofcomponentevaluations.The component evaluations are implemented by projecting out the individual components from the submitted KB, such that each component output file isformattedinthesamewayasasubmissiontotheKBPtrackforthatcomponent, and evaluating each output file as though it had been submitted directly to the standalone track for that component.
ExceptfortheSFtask,allcomponenttasksareevaluatedusinggoldstandardannotationsonacommonsetofapproximately500"core"documents.
ComponentEvaluations
ThefollowingcomponentfilesareprojectedfromeachColdStartKBforthecomponentevaluations:
1. Entitydiscoveryandlinking(EDL):AnEDLfileconsistingofnameandnominalmentionsandlinksforPER,ORG,GPE,FAC,andLOCentitiesfromthe"core"documentsusedtoevaluatesubmissionstotheEDLtrack.LinkscanbetoeitheranodeinthereferenceKB(TACKBPKnowledgeBaseII-BaseKB)or(iftheentitydoesnotexistinthereferenceKB)a
22
NILnodecorrespondingtoanentitynodeinthesubmittedKB.EvaluationoftheEntityDiscoveryandLinkingcomponentofsubmittedColdStartKBswillbeidenticaltoscoringforthe2017TACTrilingualEntityDiscoveryandLinkingtask.PleaseseeTACKBP2017EntityDiscoveryandLinkingTaskDescriptionforcompletedetailsonscoring.
2. SlotFilling:AnSFfileconsistingofslotfillersandjustificationsfoundintheKBbyapplyingColdStartevaluationqueriesthatinvolveonlySFpredicates.ThecomponentSFevaluationisidenticaltothecompositeKBevaluation,exceptthattheSFevaluationincludesonlyCSqueriesthatinvolveonlySFslots.BecauseSFsystemsareallowedtosubmitonlyonejustificationperrelation,onlythetoprankedjustificationperrelationwillbeconsideredfortheSFcomponentevaluationofKBs.
3. EventNuggetDetectionandCoreference:AnENfileconsistingofeventmentionsandwithin-documentcoreferencefromthe"core"documents.EvaluationoftheEventNuggetcomponentofsubmittedColdStartKBswillbeidenticaltoscoringforthe2017TACEventNuggetDetectionandCoreferencetask.PleaseseeTACKBP2017EventNuggetDetectionandCoreferenceTaskDescriptionforcompletedetailsonscoring.
4. EventArgumentandLinking:Asetof"arguments"files,eachfileconsistingofeventargumentassertions(includingjustifications)froma"core"document;asetof"linking"files,eachfileconsistingofcoreferenceofassertionsinthecorresponding"arguments"file.EvaluationoftheEventArgumentcomponentofsubmittedColdStartKBswillbeidenticaltoscoringforthe2017TACEventArgumentandLinkingtask.PleaseseeTACKBP2017EventArgumentandLinkingTaskDescriptionforcompletedetailsonscoring.
5. Sentiment:AsetofpredictedERExmlfiles,eachfileconsistingofname,nominal,andpronominalmentionsandcoreferenceforPER,ORG,GPE,FAC,andLOCentitiesfroma"core"document;asetofBeStxmlfiles,eachfileconsistingofsentiment(includingprovenance)fromasourcetowardsatargetentityinthecorrespondingpredictedEREfile.EvaluationofthesentimentcomponentofsubmittedColdStartKBswillbeidenticaltoscoringforthe2017TACBeliefandSentimenttask,exceptthatonlysentimenttowardsentitieswillbeevaluated.PleaseseeTACKBP2017BeliefandSentimentTaskDescriptionforcompletedetailsonscoring.
CompositeEvaluationAssessment
ColdStart2017assessmentandscoringwillproceedasfollows:Theresponsesforeachevaluationquery(frombothCSKBandCSSFsystemsandfromhuman-generatedresults)willbepooled,andeachresponsewillbeassessedbyaperson.TheresultoffollowingthefirstrelationwillbeassessedasifitwereaSlotFillingquery.ThesecondrelationinthequerywillalsobeassessedasaSlotFillingquery,butonlyifthefillforthefirstrelationiscorrect.Ifthefillforthefirstrelationisnotcorrect,eachfillforthesecondrelationisautomaticallycountedasWrong.Forexample,ifthequeryasksfortheagesofthesiblingsof“BartSimpson,”andthesubmittedknowledgebasegives“Lisaage8”and“Milhouseage10”assiblings,thenonlythereportedageofLisawillbeassessed(MilhouseisnotBart’ssibling),andthereportedageofMillhousewillautomaticallybecountedasWrong.
ColdStartusespseudo-slotscoringtoevaluatemultiple-hopqueries,inwhicheachevaluationqueryistreatedasifitselectsasingleindivisibleslot.Forexample,anevaluationquerythatasksforthechildrenofthesiblingsofanentitywillbescoredasifitwereaqueryaboutavirtual
23
per:nieces_and_nephewsslot.15TheguidelinesinTACKBP2015SlotDescriptionsspecifywhethereachofthecomponentslotsofapseudo-slotissingle-valued(e.g.,per:date_of_birth)orlist-valued(e.g.,per:employee_of,per:children).Apseudoslotissingle-valuedifeachofitscomponentslotsissingle-valued,andlist-valuedotherwise.IncontrasttotheSlotFillingtask,ColdStartKBsubmissionsmaycontainmultiplefillsforsingle-valuedslots.Ifsucharepresentinthesubmission,LDCwillassesstheslotfillwiththehighestconfidencevalue,andwillassessadditionalslotfillsifresourcesallow.Ifmorethanonesuchslotfillsharesthesameconfidencevalue,theslotfillthatappearsearlierintheoutputwillbeconsideredtohavehigherconfidence.
EachCSSFslotfillerresponse(orCSKBobjectofeachcomponentrelationthatmakesupasingleevaluationqueryresponse)isassessedasCorrect,ineXact,orWrong.Aresponseisinexactifiteitherincludesonlyapartofthecorrectanswerorincludesthecorrectanswerplusextraneousmaterial.InexactanswersarecountedasWrongforthepurposesofscoring.Iftherelationobjectisanevent,thecanonical_mentionshouldbeanEREeventtriggerbut(giventhedifficultyofdeterminingexactextentsofeventmentions),assessorsin2017willbelenientwhenassessingtheextentofanotherwisecorrecteventmention.
Foreachquery,allsystemresponsesinwhichtheslotfillerisassessedasCorrectorineXactwillbepartitionedintoequivalenceclasses,whereslotfillersinthesameequivalenceclassrepresentthesameentity,event,orvalue(asinthecaseofdates).EachCorrectorineXactresponsewillreceiveanannotationforfillermentiontype(eitherNAMorNOM),andeachequivalenceclasswillreceiveanannotationforequivalenceclassmentiontype(NAMiftheassessorcanfindanamedmentionforthefilleranywhereintheprovenancesinanyoftheresponses;otherwise,NOMifonlynominalmentionsappearintheprovenancesofallresponses).
Pseudo-slotswillbescoredjustasslotsintheSlotFillingtask,withtheadditionalconstraintthatboththeslotfillandthepathleadingtothatfillmustbecorrectfortheentiretytobejudgedcorrect.ToreceivecreditforidentifyingMaggieSimpsonasPattyBouvier’sniece,theknowledgebasemustnotonlyincludeMaggieastheslotfill,butmustalsorepresentMaggieasMarge’schild,andMargeasPatty’ssibling:16
Evaluationquery: NiecesandnephewsofPattyBouvier(per:siblings,per:children) GroundTruth: :PattyBouvier per:siblings :MargeSimpson :MargeSimpson per:children :MaggieSimpson Submission: :PattyBouvier per:siblings :MargeSimpson
:MargeSimpson per:children :MaggieSimpsonÞcorrect
AKBthatindicatedthatMaggiewasPatty’sniecebecauseshewasPatty’ssisterSelma’schildwouldbescoredasincorrect:
Evaluationquery: NiecesandnephewsofPattyBouvier(per:siblings,per:children) GroundTruth: :PattyBouvier per:siblings :MargeSimpson :MargeSimpson per:children :MaggieSimpson Submission: :PattyBouvier per:siblings :SelmaBouvier
:SelmaBouvier per:children :MaggieSimpsonÞincorrect
15Apseudo-slotissimilartotheconceptofarolechain,whichissupportedbysomeknowledgerepresentationsystemsbasedondescriptionlogic,includingOWL2.16Ineachoftheseexamples,onlythesubject,predicateandobjectareshown,andonlyasubsetoftherelevantknowledgebaseispresented.Eachentityisnamedafterthementionthatgaverisetoit.
24
Inaddition,theobjectofthefinalrelationinapseudo-slotmayberatedasredundantifitisequivalenttoanotherfillforthepseudo-slot.RedundantanswersarecountedasWrongforthepurposesofscoring:
Evaluationquery: NiecesandnephewsofPattyBouvier(per:siblings,per:children)GroundTruth: :PattyBouvier per:siblings :MargeSimpson
:MargeSimpson per:children :MaggieSimpson :MaggieSimpson per:alternate_names "Margaret Simpson" Submission: :PattyBouvier per:siblings :MargeSimpson :MargeSimpson per:children :MaggieSimpsonÞcorrect
:MargeSimpson per:children :MargaretSimpsonÞredundant
However,objectsofrelationsotherthanthefinalrelationwillneverberatedasredundant:
Evaluationquery: NiecesandnephewsofPattyBouvier(per:siblings,per:children)GroundTruth: :PattyBouvier per:siblings :MargeSimpson
:MargeSimpson per:children :LisaSimpson :MargeSimpson per:children :BartSimpson :MargeSimpson per:alternate_names "Marjorie Simpson" Submission: :PattyBouvier per:siblings :MargeSimpson :PattyBouvier per:siblings :MarjorieSimpson :MargeSimpson per:children :LisaSimpsonÞcorrect
:MarjorieSimpson per:children :BartSimpsonÞcorrectHere,MargeSimpsonandMarjorieSimpsonrepresentthesamepersoninthegroundtruth,buttwodistinctentitiesintheKB.However,becausethequeryisaboutMarge’schildrenandnotaboutMargeherself,bothresponsestotheevaluationqueryareassessedascorrect.
SinceinColdStartthefactsbeingevaluatedcomefromsequencesoftriples,confidencescoreswouldneedtobecombinedifwewantedtogenerateconfidencescoresforaderivedpseudo-relation.Threegeneralscorecombinationfunctionsaremin,maxandproduct.Theproperwaytocombinescoresofcoursedependsonthemeaningofthosescores;fornow,ColdStartwillselecttheproductfunctionasareasonableconfidencecombinationfunction.
CompositeEvaluationScoring
Giventheaboveapproachtoassessment,basicscoringforagivensystemproceedsasfollows:
• EachresponseassessedasWrongorineXact,iscountedasSpurious• EachresponseforRound2whoseRound1parentfillerisassessedasWrongorineXact,is
countedasSpurious.ThisscoringpolicyassumesthattheColdStartsystemoutputisintendedforfullyautomaticdownstreamanalytics;hence,ahop1responseisWrongifthehop0parentresponseisWrong.
• ResponsesassessedasCorrectorInexactaregroupedintoequivalenceclasses.Forquerieswithanentityasthepredicateobject,ifthesystemhasaNAMentitymentionintheequivalenceclass,orifthesystemhasonlyNOMentitymentionsandtheequivalenceclassisannotatedasNOM,thentheresponsesintheequivalenceclassarecountedasRight;otherwise,ifthesystemhasonlyNOMentitymentionsintheequivalenceclassandtheequivalenceclassisannotatedasNAM,thentheresponsesintheequivalenceclassarecountedasIgnore(i.e.,treatedasifitwasneverreturnedbythesystem)andremovedfromtheequivalenceclass.Thus,namedentitymentionsarepreferredandanamedmentionmustbereturnedifoneexists.
25
• Reference=numberofsingle-valuedpseudo-slotswithacorrectresponse+numberofequivalenceclasses17foralllist-valuedpseudo-slots
Aftereachpooledresponsehasbeenassessedand(ifassessedasCorrectorInexact)putintoanequivalenceclass,NISTwillscoreasubmittedKBbycomputingavariantofmeanaverageprecision(MAP)fortheKB.Averageprecision(AP)foragivenqueryandsubmittedrun,iscomputedinthefollowingway:
0.LetkbethemaximumnumberofjustificationsassessedperrelationassertedintheKB.ForTAC2017,itisexpectedthatk=3forKBsubmissions.(SFsubmissionswillhavek=1.)Wewillalsorequirethatforagivenfiller,atmostonejustificationisreturnedperdocument(ResolveQuerieswillconsideronlythehighestconfidencejustificationforthatdocument).
1.Foreachcandidatefiller(node)returnedforthequery,letthenodeconfidencebetheaggregateoftheconfidencesofitsjustifications--uptokjustificationsj1,j2,...jkpernode,havingconfidencec1,c2,..ck.Ifthenodehasaparent(i.e.,thefillerisahop_1filler),itsnodeconfidenceistheproductofthenodeconfidenceofitsparententity,andtheaggregateoftheconfidencesofitsjustifications.Therearemanypossiblewaysofaggregatingjustification-levelconfidencevaluestoproduceanode-levelconfidencevalue,andwhichaggregationfunctionwilldependlargelyonhowconfidencevaluesareusedineachusecase.ForTAC2017,theconfidenceofanodewithjustificationshavingconfidence(c1,c2,…,ck),assumingtheconfidencesaresortingindecreasingorder,willbeanormalizedweightedsumoftheconfidencevalues,weightedbasedontherankofthejustification:
(c1/1+c2/2+….+ck/k)/(1+½+…+1/k).
2.Rankthecandidatefillersbytheirnodeconfidence.
3.Godownthelistofnodes,andassignavaluevtoeachnode,0
26
thefilleriszeroifthefillercannotbematchedtoanequivalenceclass,eitherbecausealljustificationsinthefillerareWrong,orbecauseallequivalenceclassesinthefillerhavealreadybeenmatchedtosomeotherhigherrankedfiller(s).ThevaluevmeasuresthenumberofcorrectjustificationsdocumentsthattheKBprovidesforthematchedequivalenceclass;visarecallmeasureoverjustificationdocuments.
4.WethencomputeAPasusual,exceptthata"Correct"itemintherankedlistdoesnotalwaysgetcountedas"1",buthassomevalue0
27
3. MonolingualChinese:entitymentions,slotfillsandprovenancesareextractedonlyfromChinesedocuments.EvaluationquerieswillcontainentrypointsonlyfromChinesedocuments.
4. Cross-lingual:entitymentions,slotfillsandprovenancesareextractedfromanycombinationofEnglish,Spanish,andChinesedocuments.Evaluationquerieswillcontainentrypointsfromanyoftheselanguages,andtheslotfillerandjustificationscancomefromalanguagedifferentfromtheentrypoint.Becausejustificationspansforasinglejustificationmustcomefromthesamedocument,thecross-lingualnatureofthisevaluationconditionisduetotheexerciseofcross-lingualEDL.
IfateamsubmitsaruninvolvingmorethanonelanguageundertheCross-lingualcondition,itmustalsosubmitatleastonerununderthemonolingualconditionforeachlanguageinvolved(withadescriptionofwhichmonolingualrunconfigurationswereusedforeachcross-lingualrun).
Submittedrunsmustberanked(1-5).TherunIDincludedineachteam'ssubmissionfilemustbeaconcatenationoftheteam'sTACKBP2017teamID,thetask(KBorSF),thelanguagecondition(ENG,CMN,SPA,orXLING),andarank(1-5);thus"Acme_KB_XLING_1"wouldbethetop-rankedrunfortheAcmeteamfortheCSKBtaskunderthecross-lingualcondition.
Thetop-rankedsubmissionmustbemadeasa‘closed’system;inparticular,itmustnotaccesstheWebduringtheevaluationperiod.Allsubmissionsmustobeythefollowingexternalresourcerestrictions:
• Structuredknowledgebases(e.g.,Wikipediainfoboxes,DBPedia,Freebase)maynotbeusedtodirectlyfillslotsordirectlyvalidatecandidateslotfillers.
• Structuredknowledgebaseentriesfortargetentitiesmaynotbeedited,eitherduring,oraftertheevaluation.
Inaddition,becauseColdStartfocusesontheconditionwheretheknowledgebaseisinitiallyempty,weaskthateachparticipatingsitesubmitatleastonerunthatconsultsexternalentityknowledgebasesonlyafterentitiesandrelationshavebeenextractedfromthedocumentcollection.Detailsaboutsubmissionprocedureswillbecommunicatedtothetrackmailinglist.ToolstovalidateformatswillbeavailableontheTACWebsite(http://www.nist.gov/tac/2016/KBP/ColdStart/tools.html).
Appendix
28
Relation Inverse(s)per:children per:parents per:other_family per:other_family per:parents per:children per:siblings per:siblings per:spouse per:spouse per:employee_or_member_of {org,gpe}:employees_or_members* per:schools_attended org:students* per:city_of_birth gpe:births_in_city* per:stateorprovince_of_birth gpe:births_in_stateorprovince* per:country_of_birth gpe:births_in_country* per:cities_of_residence gpe:residents_of_city* per:statesorprovinces_of_residence gpe:residents_of_stateorprovince per:countries_of_residence gpe:residents_of_country* per:city_of_death gpe:deaths_in_city* per:stateorprovince_of_death gpe:deaths_in_stateorprovince* per:country_of_death gpe:deaths_in_country* org:shareholders {per,org,gpe}:holds_shares_in* org:founded_by {per,org,gpe}:organizations_founded* org:top_members_employees per:top_member_employee_of* {org,gpe}:member_of org:members org:members {org,gpe}:member_of org:parents {org,gpe}:subsidiaries org:subsidiaries org:parents org:city_of_headquarters gpe:headquarters_in_city* org:stateorprovince_of_headquarters gpe:headquarters_in_stateorprovince* org:country_of_headquarters gpe:headquarters_in_country*
Table2.Entity-valuedSFslots.SlotswithasterisksrepresentinverserelationsthatwillneedtobeaddedbyparticipantsfrompreviousyearsSlotFillingtask(2014andearlier).Thetypequalifierofeachrelation(per,orgorgpe)isthetypeofitssubject,whilethetypequalifierforitsinverseisthetypeofitsobject.Asetoftypesmeansthatanyofthosetypesisacceptableforthatslot.Allsubmittedslotnamesmustuseonlyasingletypespecification.
per:alternate_names org:alternate_names per:date_of_birth org:political_religious_affiliation per:age org:number_of_employees_members per:origin org:date_founded per:date_of_death org:date_dissolved per:cause_of_death org:website per:title per:religion per:charges
Table3.String-valuedSFslots.
29
SentimentPredicates
Subject Predicate Object InversePredicate
PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:dislikes PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:is_disliked_by
PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:is_disliked_by PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:dislikes
PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:is_liked_by PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:likes
PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:likes PER,ORG,GPE,LOC,FAC
{per,org,gpe,loc,fac}:is_liked_by
30
EventPredicates(eventassubject)
SubjectPredicate Object InversePredicate
CONFLICT.ATTACKconflict.attack:attacker
PER,ORG,GPE,STRING
{per,gpe,org}:conflict.attack_attacker
CONFLICT.ATTACKconflict.attack:instrument STRING none
CONFLICT.ATTACKconflict.attack:target
PER,ORG,GPE,FAC,STRING
{per,gpe,org,fac}:conflict.attack_target
CONFLICT.ATTACKconflict.attack:time STRING none
CONFLICT.ATTACKconflict.attack:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:conflict.attack_place
CONFLICT.DEMONSTRATE
conflict.demonstrate:entity PER,ORG,STRING {per,org}:conflict.demonstrate_entity
CONFLICT.DEMONSTRATE
conflict.demonstrate:time STRING none
CONFLICT.DEMONSTRATE
conflict.demonstrate:place GPE,LOC,FAC,STRING
{gpe,loc,fac}:conflict.demonstrate_place
CONTACT.BROADCASTcontact.broadcast:audience
PER,ORG,GPE,STRING
{per,org,gpe}:contact.broadcast_audience
CONTACT.BROADCASTcontact.broadcast:entity
PER,ORG,GPE,STRING
{per,org,gpe}:contact.broadcast_entity
CONTACT.BROADCASTcontact.broadcast:time STRING none
CONTACT.BROADCASTcontact.broadcast:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:contact.broadcast_place
CONTACT.CONTACTcontact.contact:entity
PER,ORG,GPE,STRING
{per,org,gpe}:contact.contact_entity
CONTACT.CONTACTcontact.contact:time STRING none
CONTACT.CONTACTcontact.contact:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:contact.contact_place
CONTACT.CORRESPONDENCE
contact.correspondence:entity
PER,ORG,GPE,STRING
{per,org,gpe}:contact.correspondence_entity
CONTACT.CORRESPONDENCE
contact.correspondence:time
STRING none
CONTACT.CORRESPONDENCE
contact.correspondence:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:contact.correspondence_place
CONTACT.MEETcontact.meet:entity
PER,ORG,GPE,STRING
{per,org,gpe}:contact.meet_entity
CONTACT.MEETcontact.meet:time STRING none
CONTACT.MEETcontact.meet:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:contact.meet_place
JUSTICE.ARREST-JAILjustice.arrest-jail:agent
PER,ORG,GPE,STRING
{per,org,gpe}:justice.arrest-jail_agent
31
JUSTICE.ARREST-JAILjustice.arrest-jail:crime STRING none
JUSTICE.ARREST-JAILjustice.arrest-jail:person PER,STRING per:justice.arrest-jail_person
JUSTICE.ARREST-JAILjustice.arrest-jail:time STRING none
JUSTICE.ARREST-JAILjustice.arrest-jail:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:justice.arrest-jail_place
LIFE.DIElife.die:agent
PER,ORG,GPE,STRING
{per,org,gpe}:life.die_agent
LIFE.DIElife.die:instrument STRING none
LIFE.DIElife.die:victim PER,STRING per:life.die_victim
LIFE.DIElife.die:time STRING none
LIFE.DIElife.die:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:life.die_place
LIFE.INJURElife.injure:agent
PER,ORG,GPE,STRING
{per,org,gpe}:life.injure_agent
LIFE.INJURElife.injure:instrument STRING none
LIFE.INJURElife.injure:victim PER,STRING per:life.injure_victim
LIFE.INJURElife.injure:time STRING none
LIFE.INJURElife.injure:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:life.injure_place
MANUFACTURE.ARTIFACT
manufacture.artifact:agent PER,ORG,GPE,STRING
{per,org,gpe}:manufacture.artifact_agent
MANUFACTURE.ARTIFACT
manufacture.artifact:artifact
FAC,STRING fac:manufacture.artifact_artifact
MANUFACTURE.ARTIFACT
manufacture.artifact:instrument
STRING none
MANUFACTURE.ARTIFACT
manufacture.artifact:time STRING none
MANUFACTURE.ARTIFACT
manufacture.artifact:place GPE,LOC,FAC,STRING
gpe,loc,fac}:manufacture.artifact_place
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:agent
PER,ORG,GPE,STRING
{per,org,gpe}:movement.transport-artifact_agent
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:artifact
FAC,STRING fac:movement.transport-artifact_artifact
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:destination
GPE,LOC,FAC,STRING
{gpe,loc,fac}:movement.transport-artifact_destination
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:instrument
STRING none
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:origin
GPE,LOC,FAC,STRING
{gpe,loc,fac}:movement.transport-artifact_origin
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:time
STRING none
MOVEMENT.TRANSPORmovement.transport-person:agent
PER,ORG,GPE,STRING
{per,org,gpe}:movement.transport-
32
T-PERSON person_agent
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:destination
GPE,LOC,FAC,STRING
{gpe,loc,fac}:movement.transport-person_destination
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:instrument
STRING none
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:origin
GPE,LOC,FAC,STRING
{gpe,loc,fac}:movement.transport-person_origin
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:person
PER,STRING per:movement.transport-person_person
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:time
STRING none
PERSONNEL.ELECTpersonnel.elect:agent PER,ORG,GPE,STRIN
G{per,org,gpe}:personnel.elect_agent
PERSONNEL.ELECTpersonnel.elect:person PER,STRING per:personnel.elect_person
PERSONNEL.ELECTpersonnel.elect:position STRING none
PERSONNEL.ELECTpersonnel.elect:time STRING none
PERSONNEL.ELECTpersonnel.elect:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:personnel.elect_place
PERSONNEL.END-POSITION
personnel.end-position:entity
ORG,GPE,STRING {org,gpe}:personnel.end-position_entity
PERSONNEL.END-POSITION
personnel.end-position:person
PER,STRING per:personnel.end-position_person
PERSONNEL.END-POSITION
personnel.end-position:position
STRING none
PERSONNEL.END-POSITION
personnel.end-position:time
STRING none
PERSONNEL.END-POSITION
personnel.end-position:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:personnel.end-position_place
PERSONNEL.START-POSITION
personnel.start-position:entity
ORG,GPE,STRING {org,gpe}:personnel.start-position_entity
PERSONNEL.START-POSITION
personnel.start-position:person
PER,STRING per:personnel.start-position_person
PERSONNEL.START-POSITION
personnel.start-position:position
STRING none
PERSONNEL.START-POSITION
personnel.start-position:time
STRING none
PERSONNEL.START-POSITION
personnel.start-position:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:personnel.start-position_place
TRANSACTION.TRANSACTION
transaction.transaction:beneficiary
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transaction_beneficiary
TRANSACTION.TRANSACTION
transaction.transaction:giver
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transaction_giver
TRANSACTION.TRANSACTION
transaction.transaction:recipient
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transaction_recipient
TRANSACTION.TRANSACTION
transaction.transaction:time
STRING none
33
EventPredicates(eventasobject)
SubjectPredicate Object InversePredicate
PER,ORG,GPE{per,org,gpe}:conflict.attack_attacker CONFLICT.ATTACK conflict.attack:attacker
GPE,LOC,FAC{gpe,loc,fac}:conflict.attack_place CONFLICT.ATTACK conflict.attack:place
PER,ORG,GPE,FAC
{per,org,gpe,fac}:conflict.attack_target CONFLICT.ATTACK conflict.attack:target
PER,ORGper,org:conflict.demonstrate_entity
CONFLICT.DEMONSTRATE
conflict.demonstrate:entity
TRANSACTION.TRANSACTION
transaction.transaction:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:transaction.transaction_place
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:beneficiary
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transfer-money_beneficiary
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:giver
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transfer-money_giver
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:money
STRING none
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:recipient
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transfer-money_recipient
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:time
STRING none
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:transaction.transfer-money_place
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:beneficiary
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transfer-ownership_beneficiary
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:giver
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transfer-ownership_giver
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:recipient
PER,ORG,GPE,STRING
{per,org,gpe}:transaction.transfer-ownership_recipient
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:thing
FAC,ORG,STRING {fac,org}:transaction.transfer-ownership_thing
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:time
STRING none
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:place
GPE,LOC,FAC,STRING
{gpe,loc,fac}:transaction.transfer-ownership_place
34
GPE,LOC,FAC{gpe,loc,fac}:conflict.demonstrate_place
CONFLICT.DEMONSTRATE
conflict.demonstrate:place
PER,ORG,GPE{per,org,gpe}:contact.broadcast_audience
CONTACT.BROADCAST contact.broadcast:audience
PER,ORG,GPE{per,org,gpe}:contact.broadcast_entity CONTACT.BROADCAST contact.broadcast:entity
GPE,LOC,FAC{gpe,loc,fac}:contact.broadcast_place CONTACT.BROADCAST contact.broadcast:place
PER,ORG,GPE{per,org,gpe}:contact.contact_entity CONTACT.CONTACT contact.contact:entity
GPE,LOC,FAC{gpe,loc,fac}:contact.contact_place CONTACT.CONTACT contact.contact:place
PER,ORG,GPE{per,org,gpe}:contact.correspondence_entity
CONTACT.CORRESPONDENCE
contact.correspondence:entity
GPE,LOC,FAC{gpe,loc,fac}:contact.correspondence_place
CONTACT.CORRESPONDENCE
contact.correspondence:place
PER,ORG,GPE{per,org,gpe}:contact.meet_entity CONTACT.MEET contact.meet:entity
GPE,LOC,FAC{gpe,loc,fac}:contact.meet_place CONTACT.MEET contact.meet:place
PER,ORG,GPE{per,org,gpe}:justice.arrest-jail_agent JUSTICE.ARREST-JAIL justice.arrest-jail:agent
PERper:justice.arrest-jail_person JUSTICE.ARREST-JAIL justice.arrest-jail:person
GPE,LOC,FAC{gpe,loc,fac}:justice.arrest-jail_place JUSTICE.ARREST-JAIL justice.arrest-jail:place
PER,ORG,GPE{per,org,gpe}:life.die_agent LIFE.DIE life.die:agent
GPE,LOC,FAC{gpe,loc,fac}:life.die_place LIFE.DIE life.die:place
PERper:life.die_victim LIFE.DIE life.die:victim
PER,ORG,GPE{per,org,gpe}:life.injure_agent LIFE.INJURE life.injure:agent
GPE,LOC,FAC{gpe,loc,fac}:life.injure_place LIFE.INJURE life.injure:place
PERper:life.injure_victim LIFE.INJURE life.injure:victim
PER,ORG,GPE{per,org,gpe}:manufacture.artifact_agent MANUFACTURE.ARTIFAC
Tmanufacture.artifact:agent
FACfac:manufacture.artifact_artifact
MANUFACTURE.ARTIFACT
manufacture.artifact:artifact
GPE,LOC,FAC{gpe,loc,fac}:manufacture.artifact_place
MANUFACTURE.ARTIFACT
manufacture.artifact:place
PER,ORG,GPE{per,org,gpe}:movement.transport-artifact_agent
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:agent
FACfac:movement.transport-artifact_artifact
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:artifact
GPE,LOC,FAC{gpe,loc,fac}:movement.transport-artifact_destination
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:destination
GPE,LOC,FAC{gpe,loc,fac}:movement.transport-artifact_origin
MOVEMENT.TRANSPORT-ARTIFACT
movement.transport-artifact:origin
PER,ORG,GPE{per,org,gpe}:movement.transport-person_agent
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:agent
GPE,LOC,FAC{gpe,loc,fac}:movement.transport- movement.transport-
35
person_destination MOVEMENT.TRANSPORT-PERSON
person:destination
GPE,LOC,FAC{gpe,loc,fac}:movement.transport-person_origin
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:origin
PERper:movement.transport-person_person
MOVEMENT.TRANSPORT-PERSON
movement.transport-person:person
PER,ORG,GPE{per,org,gpe}:personnel.elect_agent PERSONNEL.ELECT personnel.elect:agent
PERper:personnel.elect_person PERSONNEL.ELECT personnel.elect:person
GPE,LOC,FAC{gpe,loc,fac}:personnel.elect_place PERSONNEL.ELECT personnel.elect:place
ORG,GPE{org,gpe}:personnel.end-position_entity PERSONNEL.END-
POSITIONpersonnel.end-position:entity
PERper:personnel.end-position_person PERSONNEL.END-
POSITIONpersonnel.end-position:person
GPE,LOC,FAC{gpe,loc,fac}:personnel.end-position_place
PERSONNEL.END-POSITION
personnel.end-position:place
ORG,GPE{org,gpe}:personnel.start-position_entity PERSONNEL.START-
POSITIONpersonnel.start-position:entity
PERper:personnel.start-position_person PERSONNEL.START-
POSITIONpersonnel.start-position:person
GPE,LOC,FAC{gpe,loc,fac}:personnel.start-position_place
PERSONNEL.START-POSITION
personnel.start-position:place
PER,ORG,GPE{per,org,gpe}:transaction.transaction_beneficiary
TRANSACTION.TRANSACTION
transaction.transaction:beneficiary
PER,ORG,GPE{per,org,gpe}:transaction.transaction_giver
TRANSACTION.TRANSACTION
transaction.transaction:giver
GPE,LOC,FAC{gpe,loc,fac}:transaction.transaction_place
TRANSACTION.TRANSACTION
transaction.transaction:place
PER,ORG,GPE{per,org,gpe}:transaction.transaction_recipient
TRANSACTION.TRANSACTION
transaction.transaction:recipient
PER,ORG,GPE{per,org,gpe}:transaction.transfer-money_beneficiary
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:beneficiary
PER,ORG,GPE{per,org,gpe}:transaction.transfer-money_giver
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:giver
GPE,LOC,FAC{gpe,loc,fac}:transaction.transfer-money_place
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:place
PER,ORG,GPE{per,org,gpe}:transaction.transfer-money_recipient
TRANSACTION.TRANSFER-MONEY
transaction.transfer-money:recipient
PER,ORG,GPE{per,org,gpe}:transaction.transfer-ownership_beneficiary
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:beneficiary
PER,ORG,GPE{per,org,gpe}:transaction.transfer-ownership_giver
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:giver
GPE,LOC,FAC{gpe,loc,fac}:transaction.transfer- transaction.transfer-
36
ownership_place TRANSACTION.TRANSFER-OWNERSHIP
ownership:place
PER,ORG,GPE{per,org,gpe}:transaction.transfer-ownership_recipient
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:recipient
FAC,ORG{fac,org}:transaction.transfer-ownership_thing
TRANSACTION.TRANSFER-OWNERSHIP
transaction.transfer-ownership:thing
ChangeHistory• Version1.0
o Originaldraftversion,basedonthe2016specification