2015 ESWC FrameBase presentation

  • Published on
    10-Feb-2017

  • View
    46

  • Download
    2

Transcript

FrameBase: Representing N-ary Relations Using Semantic FramesJacobo RoucesAalborg Universityjrg@es.aau.dkGerard de MeloTsinghua Universitygdm@demelo.orgKatja HoseAalborg Universitykhose@cs.aau.dkJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 2Ways to represent N-ary relations. Using Direct Binary Relations Used by default in most KBs. Dereified. RDF reification YAGO,YAGO2s Subproperties Proposed in [Nguyen et al, WWW 2014] Neo-davidsonian representations To an extent used in most Kbs that include events. Freebase, FramebaseJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 3Ways to represent N-ary relationsDirect Binary Relations Pairwise properties around an event (unreified) From N up to N(N-1) triples:person1 gotMarriedWith person2person1 gotMarriedInPlace placeperson2 gotMarriedInPlace placeperson1 gotMarriedOnDate timeperson2 gotMarriedOnDate timeperson1 ceremonyType marriageCeremonyTypeperson2 ceremonyType marriageCeremonyTypeplace holdWeddingOnDate time Without events, connections are unknown: Sarkozy gotMarriedWith Carla_BruniSarkozy gotMarriedWith Ccilia_AttiasSarkozy gotMarriedOnDate 2007Sarkozy gotMarriedOnDate 1996?Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 4e1 e2e3Ways to represent N-ary relations.Direct Binary Relationse1 p e2 .e2 q e3 .e3 r e4 .Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 5Ways to represent N-ary relationsRDF reificatione1Statemente2 Original triplee1 p e2 Reified with additional triples. r signifies the triple:r rdf:type rdf:Statementr rdf:subject e1r rdf:property pr rdf:object e2 RDF reification is different from (general) reification, where the new entity r would signify, not a triple but the event or frame evoked by a property. This other kind is central to FrameBase, and will come later.Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 6Ways to represent N-ary relationsRDF reificatione1 e2e3Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 7Ways to represent N-ary relationsRDF reificatione1Statemente2e3StatementStatementJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 8Ways to represent N-ary relationsRDF reification Possible third way: reifying a primary triple (YAGO). But: 4-fold overhead when using pure RDF, or need for quads. Lower triplestore performance and cumbersome queries. The advantage (including also the direct binary relation) is only for the primary pair. For the other direct binary relations, more reifications are needed. Which one is the primary pair? Can the user replicate the choice? Mixing metadata with data leads to ambiguity and errors in LOD: Something like :factId :time 2013 would mean that Einstein won the Nobel Prize in the 21st century or that the triple was created at that time? Non-unique triple ids when several instances of the event share the primary pair.Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 9Ways to represent N-ary relationsNeo-Davidsonian representation Reified properties (connecting properties around an event). N+1 triples: event type marriageevent partner Sarkozyevent partner Carla_Brunievent time 2007event location Parisevent manner civilCeremony Unlike the case with direct binary predicates, events can be separatedevent2 type Marriageevent2 partner1 Sarkozyevent2 partner2 Ccilia_Attiasevent2 time 1996A.k.a. Neo-Davidsonian representationJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 10Ways to represent N-ary relationsNeo-Davidsonian representation Example from http://purl.org/vocab/bio/0.1/Marriage_:e a bio:Marriage ; dc:date "1903" ; bio:partner dbpedia:Albert_Einstein ; bio:partner dbpedia:Mileva_Mari%C4%87 ; bio:place dbpedia:Bern .http://purl.org/vocab/bio/0.1/MarriageJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 11Ways to represent N-ary relationsNeo-Davidsonian representatione1Event typee2e3Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 12Ways to represent N-ary relations Using different representations is troublesome: Low recall when querying The user may use a different schema to model the query Alignment hindered Ontology alignment systems usually search direct equivalences between classes, properties, etc.Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 13FrameBase Core: RDFS schema to represent knowledge using neo-Davidsonian approach with a wide and extensible vocabulary of frames (events, situations, frames, eventualities) frame elements (outgoing properties representing frame-specific semantic roles) Vocabulary based on NLP resources (FrameNet+WordNet) This provides connection with natural language and semantic role labeling systems. Inference rules to provide direct binary predicates?f a :frame-Separating-partition.v ?f :fe-Separating-Whole ?s ?s :isPartitionedIntoParts ?o?f :fe-Separating-Parts ?o We will explain these points now...Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 14FrameBase:Core schemae1Frame typee2e3FRAME CLASSFRAME ELEMENT(FRAME-SPECIFIC SEMANTIC ROLES)FRAME INSTANCEJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 15FrameBase:Core schema Problems using FrameNet: Coverage is limited Some frames and FEs are too general Create micro-frames with LUs Too many near-equivalent frames now! Sparsity.We must cluster near-equivalent sensesby aligning and extending with WordNet (algorithm in the paper) Using synsets and lexical-semantic pointers we group Synonyms Near-equivalent senses Morphosemantic variations. e.g nominalizationsJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 16FrameBase:Core schema ..defect.v ..defection.n ..desert.v ..desertion.n..desertion_n_00055315 ..defect_v_02584097..abandon_v_00614057..deserter_n_10007109..deserter_n_10006842..retreat.v ..withdraw.v ..withdrawal.n..receding_n_00057486..pullback_n_00056688..withdraw_v_01994442..withdrawal_n_00053913:frame-Quitting_a_placedeserterturncoatapostateratterrecreantrenegadedesertionabandonmentdefectiondeserterdefectordefectdesertabandondesertdesolateforsakepullback recedingrecessionwithdrawretireretreatdrawbackpullbackmovebackrecedepullawaywithdrawalJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 17FrameBase:Core schema ..defect.v ..defection.n ..desert.v ..desertion.n..desertion_n_00055315 ..defect_v_02584097..abandon_v_00614057..deserter_n_10007109..deserter_n_10006842..retreat.v ..withdraw.v ..withdrawal.n..receding_n_00057486..pullback_n_00056688..withdraw_v_01994442..withdrawal_n_00053913:frame-Quitting_a_placedeserterturncoatapostateratterrecreantrenegadedesertionabandonmentdefectiondeserterdefectordefectdesertabandondesertdesolateforsakepullback recedingrecessionwithdrawretireretreatdrawbackpullbackmovebackrecedepullawaywithdrawalJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 18FrameBase:Reification-dereification rules Challenge using neo-davidsonian representation: The reification provided by frames is necessary when more than two slots/arguments are filled, but sometimes is not. Overhead querying and storing.Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 19FrameBase:Reification-dereification rules Solution in FrameBase: Two-layered structure. Create two levels of reification, and inference rules that connect them. Reified knowledge using frames and frame elements Dereified knowledge using direct binary predicates Rules are definite clauses (easy for inference engines)e1Event typee2e3?f a :frame-Separating-partition.vAND?f :fe-Separating-Whole ?sAND?f :fe-Separating-Parts ?oIFF?s ..-isPartitionedIntoParts ?oJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 20Example: Win_prize frame:frame-Win_prize-win.v ...-competitoryago:A_Einsteinyago:Nobel_Prizefe-Win_prize-competition fe-Win_prize-prize 1921^xsd:date...-time ...-explanation BEYOND TIME AND LOCATION!Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 21Example: Win_prize frame:frame-Win_prize-win.v ...-competitoryago:A_Einsteinyago:Nobel_Prizefe-Win_prize-competition fe-Win_prize-prize 1921^xsd:date...-timeyago:Photoelectric_effect...-explanation frame:Working_on-work.nfe-Working_on-agent ...-domain...-time1905^xsd:dateBEYOND TIME AND LOCATION!Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 22Example: Win_prize frame:frame-Win_prize-win.v ...-competitoryago:A_Einsteinyago:Nobel_Prizefe-Win_prize-competition fe-Win_prize-prize 1921^xsd:date...-time? ??yago:Photoelectric_effect...-explanation frame:Working_on-work.nfe-Working_on-agent ...-domain...-time1905^xsd:date ?BEYOND TIME AND LOCATION!Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 23Example: Win_prize frame:frame-Win_prize-win.v ...-competitoryago:A_Einsteinyago:Nobel_Prizefe-Win_prize-competition fe-Win_prize-prize 1921^xsd:date...-timewinsByCompetitor winsAtTime isWonAtTimeyago:Photoelectric_effect...-explanation frame:Working_on-work.nfe-Working_on-agent ...-domain...-time1905^xsd:date worksAtTimeBEYOND TIME AND LOCATION!Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 24FrameBase:Reification-dereification rules FrameBase: Two-layered structure:Create two levels of reification, and inference rules that connect them. Reified knowledge using frames and frame elements Dereified knowledge using direct binary predicates Rules are Horn clauses (good for inference engines) Around 15000 rules and direct binary predicates are created automatically. Different storage strategies are possible.?f a :frame-Separating-partition.vAND?f :fe-Separating-Whole ?sAND?f :fe-Separating-Parts ?oIFF?s ..-isPartitionedIntoParts ?oJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 25FrameBase:Integration rules Integration rules from source KBs can be created with SPARQL CONSTRUCT queries (and optionally a RDFier)CONSTRUCT { _:e a framebase:frame-People_by_jurisdiction-citizen.n . _:e framebase:fe-People_by_jurisdiction-Person ?person . _:e framebase:fe-People_by_jurisdiction-Jurisdiction ?country .} WHERE { ?person freebase:people.person.nationality ?country . } More examples in the DeRiVE 2015 paper Representing Specialized Events with FrameBaseJacobo Rouces, Gerard De Melo, Katja Hose02/06/15 26Results RDFS schema of size 250,407 triples Using FrameNet-WordNet mapping with precision = 0.789 It provides 19,376 frames with lexical labels A total of 18,357 microframes 11,939 LU-microframes 6,418 synset-microframes. Grouped into 8,145 logical clusters: sets of microframes whose elements are linked by a logical near-equivalence relation. We generate automatically 14,930 reificationdereification rules for the same number of direct binary predicates. Human-readable 86.59% 6.41% were correct.Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 27Data More information: http://framebase.org Data is open-source. License: CC-BY 4.0 International Everybody is welcome to publish their data using the FrameBase schema!The research leading to these results has received funding from the European Union Seventh Framework Programme (FP7/2007-2013) under grant agreement No. FP7-SEC-2012-312651 (ePOOLICE project). Additional funding was provided by National Basic Research Program of China Grants 2011CBA00300, 2011CBA00301, and NSFC Grants 61033001, 61361136003.http://framebase.org/Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 28Conclusion FrameBase offers a reusable, wide-range, semantically rich, natural-language-related and extensible schema for representation of n-ary relations, events, situations, processes, natural kinds, etc. (in general: frames). Two levels of representation: reified and dereified. Future work: Automatic integration of source KBs Interfacing with NL and QA (SEMAFOR).Jacobo Rouces, Gerard De Melo, Katja Hose02/06/15 29Slide 1Slide 2Slide 3Slide 4Slide 5Slide 6Slide 7Slide 8Slide 9Slide 10Slide 11Slide 12Slide 13Slide 14Slide 15Slide 16Slide 17Slide 18Slide 19Slide 20Slide 21Slide 22Slide 23Slide 24Slide 25Slide 26Slide 27Slide 28Slide 29