PowerPoint Presentation
LECTURE 13
SemanticRolecLabelling
ArkaitzcZubiaga,c19thcFebruary,c2018
2
WhatciscSemanticRolecLabelling?
ThematicRoles:
FrameNet.
PropBank.
ApproaihesctocSemanticRolecLabelling.
SeleitonalcRestriitons.
LECTUREc13:cCONTENTS
3
Semanti role labelling (SRL):cforcaccprediiatecorcverbcincac
sentenie,ctaskcofcidentfyingcthematicroles,csuihcascagent,cgoal,c
orcresult.
SEMANTICcROLEcLABELLING
Theme LocationPredicateAgent
4
SRLciancbecusefulcfor:
Questoncanswering.
Maihinectranslaton.
Informatoncextraiton.
Tocidentfycsenteniescwithcidentialcmeaning:
XYZciorporatoncboughtcthecstoik.
TheycsoldcthecstoikctocXYZciorporaton.
ThecstoikcwascboughtcbycXYZciorporaton.
SEMANTICcROLEcLABELLING
5
Givencacsentenie:
PaulcboughtcaciarcfromcSarahcforc£7,000
c
Identfycprediiate(s)c[i.e.cthecaiton,cgenerallycthecverb]:
PaulcboughtcaciarcfromcSarahcforc£7,000
c
Identfycthematicroles:
PaulcboughtcaciarcfromcSarahcforc£7,000
SEMANTICcROLEcLABELLING:cEXAMPLE
buyer seller
goods
money
6
Actypiialcsetcofcthematicroles.
THEMATICcROLES
7
Most frequently used themati roles:
AGENT:csubjeitc(aitvecsentenie),cbycXc(passivecsentenie).
EXPERIENCER:canimatecsubjeitcincaitvecsentenies.
THEME:cobjeitc(transitvecverb),csubjeitc(non-aitoncverb).
INSTRUMENT:cwithcX.
BENEFICIARY:cforcX.
THEMATICcROLES
8
Problem of themati roles:
Hard to ireate standard setcofcrolescorcformallycdefnecthem.
c
Alternatiectocthematicroles:
PropBank,cfewer,cmorecgeneralisedcsemanticroles.
FrameNet,cmorecroles,cspeiifictocacgroupcofcprediiates.
THEMATICcROLES
9
Iniludesc12 domainsc(body,ciogniton,chealth,ctme,…)
Conduited on frame-by-frame basis:
1.cihoosecsemanticframec(e.g.cCommerie_buy).
2.cdefnecthecframecandcitscframecelementsc(e.g.cBUYER,c
GOODS,cSELLER,cMONEY).
3.clistcprediiatescthatcevokecitc(buy.v,cpurihase.v,cpurihase.n).
4.cextraitcsenteniescforceaihcprediiatecfromcBritshcNatonalc
Corpus.
THEcFRAMENETcPROJECT
10
Frame Relaton:cdireitedcrelaton between two frames,ci.e.cac
Super_Framec(lesscdependent,cmorecabstrait)candcacSub_Framec
(thecmorecdependent,clesscabstrait).
c
e.g.cthecframecCommeriial_transaitonchasctwocsub-frames:
Commerie_goods_transfer
Commerie_money_transfer
FRAMENET:cFRAMEcRELATIONS
11
One million word iorpuscannotatedcwithcprediiatecargumentc
struitures.
DevelopedcatcUniversitycofcPennsilvania.
Prediiatescareclexiialisedconlycbycierbs.
Adjunitie arguments:cLoiatve,cTemporal,cManner,cCause,ceti.
PROPOSITIONcBANK:cPROPBANK
12
Arg0c=cagent
Arg1c=cdireitcobjeitc/cthemec/cpatent
Arg2c=cindireitcobjeitc/cbenefaitvec/cinstrumentc/catributec/cendc
state
Arg3c=cstartcpointc/cbenefaitvec/cinstrumentc/catribute
Arg4c=cendcpoint
PROPBANK:cARGUMENTcNUMBERS
13
PROPBANK:cADJUNCTS
14
Theciompanycboughtcacwheel-loadercfromcDresser.
Arg0:cTheciompany
Rel:cbought
Arg1:cacwheel-loader
Arg2-from:cDresser
PROPBANK:cEXAMPLE
15
TVcstatonscboughtc“Cosby”crerunscforcreiordcpriies.
Arg0:cTVcstatons
Rel:cbought
Arg1:c“Cosby”creruns
Arg3-for:creiordcpriies
PROPBANK:cEXAMPLE
16
BUY SELL PAY
Arg0:cbuyer Arg0:cseller Arg0:cbuyer
Arg1:ciommodity Arg1:ciommodity Arg1:cpriiecpaid
Arg2:cseller Arg2:cbuyer Arg2:cseller
Arg3:cpriie Arg3:cpriie Arg3:ciommodity
Arg4:cbenefiiary Arg4:cbenefiiary Arg4:cbenefiiary
PROPBANK:cFLEXIBILITY
17
FRAMENETcVScPROPBANK
18
Early SRL systemsc(1970s-80s):
parsercfollowedcbychand-writencrulescforceaihcverb.
diitonariescwithcverb-speiificiasecframes.
c
Current systems use maihine learning.
APPROACHEScTOcSEMANTICcROLEcLABELLING
19
AcSIMPLEcMODERNcALGORITHM
20
Usescparsedctreecascinput.
SEMANTICcROLEcLABELLING:cEXAMPLE
21
1.cPruning:cusecsimplecheuristisctocprunecunlikelycionsttuents.
2. Identiiaton:cacbinarycilassifiatoncofceaihcnodecascanc
argumentctocbeclabelledcorcacNONE.
3.cClassiiiaton:cac1-of-Ncilassifiatoncofcallcthecionsttuentscthatc
wereclabelledcascargumentscbycthecpreviouscstage.
3-STEPcVERSIONcOFcSRLcALGORITHM
22
Imbalaniecbetween:
positvecsamplesc(ionsttuentscthatcarecargumentscofc
prediiate).
negatvecsamplesc(ionsttuentscthatcarecnotcargumentscofc
prediiate).
Imbalanied data ian be hardcforcmanycilassifers.
Socwe prune the iery unlikely ionsttuents irst,candcthencusecac
ilassiferctocgetcridcofcthecrest.
WHYcPRUNINGcANDcIDENTIFICATIONcSTEPS?
23
Algorithm so far ilassiies eierything loiallyc–ceaihcdeiisionc
aboutcacionsttuentciscindependentcofcallcothers.
But this ian’t be right:cLotscofcglobalcorcjointcinteraitonsc
betweencarguments.
Acloialcsystemcmayciniorreitlyclabelctwocoverlappingc
ionsttuentscascarguments.
e.g.clabellingconecionsttuentcARG0cshouldcinireasecthec
probabilitycofcanothercbeingcARG1
FINALcSTAGE:cJOINTcINFERENCE
24
Reranking:
ThecfrstcstagecSRLcsystemcproduiescmultplecpossibleclabelsc
forceaihcionsttuent.
Thecseiondcstagecilassifescthecbestcglobalclabelcforcallc
ionsttuents.
Ofencacilassifercthatctakescallcthecinputscalongcwithcotherc
featuresc(sequeniescofclabels).
HOWcTOcDOcJOINTcINFERENCE
25
SRL:cNOTcJUSTcENGLISH
26
SEMANTICcROLEcLABELLING:cEXAMPLE
27
Prunecthecprediiatecandcallcothercelementscthatciniludecit.
SRLcEXAMPLE:cPRUNING
28
Binarycilassifiatoncofceaihcelement:cargumentcorcnot.
SRLcEXAMPLE:cARGUMENTcIDENTIFICATION
29
SRLcEXAMPLE:cARGUMENTcIDENTIFICATION
30
Multilasscilassifiatoncofcargumentctype.
SRLcEXAMPLE:cARGUMENTcCLASSIFICATION
31
SRLcEXAMPLE:cARGUMENTcCLASSIFICATION
32
Featuresctocilassifycarguments:
Governingcprediiate,ce.g.cissued.
Phrasectypecthatcitcis,ce.g.cNP.
Pathcfromcargumentctocprediiate.
Positon,ci.e.cbeforecorcafercprediiate.
Voiie,ci.e.caitvecorcpassive.
Headcwordcofcthecionsttuent.
SEMANTICcROLEcLABELLING:cFEATURES
33
Contentcword:clexiialisedcfeaturecthatcseleitscancinformatvec
wordcfromcthecionsttuent,cothercthancthechead.
POSctagcofctheciontentcword.
POSctagcofcthecheadcword.
NamedcEnttycilasscofctheciontentcwordc(loiaton,cperson,c
organisaton,…)
Containscnamedcenttes?cBooleancfeature.
SEMANTICcROLEcLABELLING:cMOREcFEATURES
34
Firstcwordcofcthecionsttuentc(andcitscPOSctag).
Lastcwordcofcthecionsttuentc(andcitscPOSctag).
Lefcandcrightcsiblingcionsttuentclabels.
Lefcandcrightcsiblingcheadcwordsc(andctheircPOSctags).
SEMANTICcROLEcLABELLING:cMOREcFEATURES
35
SRL:cAclevelcofcshallowcsemantiscforcrepresentngceventscandc
theircpartiipants.
Twociommoncarihiteitures,cforcvariousclanguages:
FrameNet:cframe-speiificroles.
PropBank:cProto-roles.
Currentcsystemscextraitcby:
Parsingcsentenie.
Findingcprediiates,cilassifyingcassoiiatedcarguments.
SEMANTICcROLEcLABELLING:cSUMMARY
36
Considercthectwocinterpretatonscof:
Icwantctoceatcsomewherecnearby.
c
1.cIcwantctoceatcincacnearbycplaie.
2.cItciscacplaiecnearbycthecobjeitcthatcIcwantctoceat.
c
Difiultcforcancalgorithmcifcwecexpeitcthecfollowingcpatern:
eatc+c[food],ce.g.ceatcihiiken
SELECTIONALcRESTRICTIONS
37
Seleitonalcrestriitonscarecassoiiatedcwithcsenses:
Thecrestaurantcseriescpizzas.c[servesc+cfood]
BritshcAirwayscseriescLondon.c[servesc+cplaie]
SELECTIONALcRESTRICTIONS
38
Commoncpraitiecforcseleitonalcrestriitonscisctocusecthesauric
suihcascWordNet.
c
e.g.cThe THEMEcofc“eat”cmustcbecWordNetcsynsetc{food,cnutrient}
SELECTIONALcRESTRICTIONS
39
RESOURCES
FrameNet:
http://framenet.icsi.berkeley.edu/
PropBank:
http://www.cs.rochester.edu/~gildea/PropBank/Sort/
NomBank:
http://nlp.cs.nyu.edu/meyers/NomBank.html
http://framenet.icsi.berkeley.edu/
http://www.cs.rochester.edu/~gildea/PropBank/Sort/
http://nlp.cs.nyu.edu/meyers/NomBank.html
40
RESOURCES:cSOFTWARE
SENNA:
https://ronan.collobert.com/senna/
SEMAFOR:
http://www.cs.cmu.edu/~ark/SEMAFOR/
MatePlus:
https://github.com/microth/mateplus
Open Sesame:
https://github.com/Noahs-ARK/open-sesame
https://ronan.collobert.com/senna/
http://www.cs.cmu.edu/~ark/SEMAFOR/
https://github.com/microth/mateplus
https://github.com/Noahs-ARK/open-sesame
41
ASSOCIATEDcREADING
Jurafsky, Daniel, and James H. Martin. 2009. Speech and Language
Processing: An Introduction to Natural Language Processing, Speech
Recognition, and Computational Linguistics. 3rd edition. Chapter 22.
Slide 1
Slide 2
Slide 3
Slide 4
Slide 5
Slide 6
Slide 7
Slide 8
Slide 9
Slide 10
Slide 11
Slide 12
Slide 13
Slide 14
Slide 15
Slide 16
Slide 17
Slide 18
Slide 19
Slide 20
Slide 21
Slide 22
Slide 23
Slide 24
Slide 25
Slide 26
Slide 27
Slide 28
Slide 29
Slide 30
Slide 31
Slide 32
Slide 33
Slide 34
Slide 35
Slide 36
Slide 37
Slide 38
Slide 39
Slide 40
Slide 41