43
-6-
theconcepts,andareexpectedtorecognizethemfromthewordslistedinthesynset.
Alexicalmatrix,therefore,canberepresentedfortheoreticalpurposesbya
mappingbetweenwrittenwordsandsynsets. SinceEnglishisrichinsynonyms,synsets
areoftensufficientfordifferentialpurposes. Sometimes,however,anappropriate
synonymisnotavailable,inwhichcasethepolysemycanberesolvedbyashortgloss,
e.g.,{board,(aperson’smeals,providedregularlyformoney)}canservetodifferentiate
thissenseofboardfromtheothers;itcanberegardedasasynsetwithasinglemember.
Theglossisnotintendedforuseinconstructinganewlexicalconceptbysomeonenot
alreadyfamiliarwithit,anditdiffersfromasynonyminthatitisnotusedtogainaccess
toinformationstoredinthementallexicon. Itfulfillsitspurposeifitenablestheuserof
WordNet,whoisassumedtoknowEnglish,todifferentiatethissensefromotherswith
whichitcouldbeconfused.
Synonymyis,ofcourse,alexicalrelationbetweenwordforms,butbecauseitis
assignedthiscentralroleinWordNet,anotationaldistinctionismadebetweenwords
relatedbysynonymy,whichareenclosedincurlybrackets,‘{’and‘}’,andotherlexical
relations,whichwillbeenclosedinsquarebrackets,‘[’and‘]’. Semanticrelationsare
indicatedbypointers.
WordNetisorganizedbysemanticrelations. Sinceasemanticrelationisarelation
betweenmeanings,andsincemeaningscanberepresentedbysynsets,itisnaturalto
thinkofsemanticrelationsaspointersbetweensynsets. Itischaracteristicofsemantic
relationsthattheyarereciprocated:ifthereisasemanticrelationRbetweenmeaning{x,
x¢,...}andmeaning{y,y¢,...},thenthereisalsoarelationR¢between{y,y¢,...}and
{x,x¢,...}. Forthepurposesofthepresentdiscussion,thenamesofthesemantic
relationswillserveadualrole:iftherelationbetweenthemeanings{x,x¢,...}and{y,
y¢,...}iscalledR,thenRwillalsobeusedtodesignatetherelationbetweenindividual
wordformsbelongingtothosesynsets. Itmightbelogicallytidiertointroduceseparate
termsfortherelationbetweenmeaningsandfortherelationbetweenforms,buteven
greaterconfusionmightresultfromtheintroductionofsomanynewtechnicalterms.
Thefollowingexamplesillustrate(butdonotexhaust)thekindsofrelationsusedto
createWordNet.
Synonymy
Fromwhathasalreadybeensaid,itshouldbeobviousthatthemostimportant
relationforWordNetissimilarityofmeaning,sincetheabilitytojudgethatrelation
betweenwordformsisaprerequisitefortherepresentationofmeaningsinalexical
matrix. Accordingtoonedefinition(usuallyattributedtoLeibniz)twoexpressionsare
synonymousifthesubstitutionofonefortheotherneverchangesthetruthvalueofa
sentenceinwhichthesubstitutionismade. Bythatdefinition,truesynonymsarerare,if
theyexistatall. Aweakenedversionofthisdefinitionwouldmakesynonymyrelativeto
acontext:twoexpressionsaresynonymousinalinguisticcontextCifthesubstitutionof
onefortheotherinCdoesnotalterthetruthvalue. Forexample,thesubstitutionof
plankforboardwillseldomaltertruthvaluesincarpentrycontexts,althoughthereare
othercontextsofboardwherethatsubstitutionwouldbetotallyinappropriate.
39
-7-
Notethatthedefinitionofsynonymyintermsofsubstitutabilitymakesitnecessary
topartitionWordNetintonouns,verbs,adjectives,andadverbs. Thatistosay,if
conceptsarerepresentedbysynsets,andifsynonymsmustbeinterchangeable,then
wordsindifferentsyntacticcategoriescannotbesynonyms(cannotformsynsets)
becausetheyarenotinterchangeable. Nounsexpressnominalconcepts,verbsexpress
verbalconcepts,andmodifiersprovidewaystoqualifythoseconcepts. Inotherwords,
theuseofsynsetstorepresentwordmeaningsisconsistentwithpsycholinguistic
evidencethatnouns,verbs,andmodifiersareorganizedindependentlyinsemantic
memory. Anargumentmightbemadeinfavorofstillfurtherpartitions:somewordsin
thesamesyntacticcategory(particularlyverbs)expressverysimilarconcepts,yetcannot
beinterchangedwithoutmakingthesentenceungrammatical.
Thedefinitionofsynonymyintermsoftruthvaluesseemstomakesynonymya
discretematter:twowordseitheraresynonymsortheyarenot. Butassome
philosophershaveargued,andmostpsychologistsacceptwithoutconsideringthe
alternative,synonymyisbestthoughtofasoneendofacontinuumalongwhich
similarityofmeaningcanbegraded. Itisprobablythecasethatsemanticallysimilar
wordscanbeinterchangedinmorecontextsthancansemanticallydissimilarwords. But
theimportantpointhereisthattheoriesoflexicalsemanticsdonotdependontruth-
functionalconceptionsofmeaning;semanticsimilarityissufficient. Itisconvenientto
assumethattherelationissymmetric:ifxissimilartoy,thenyisequallysimilartox.
Thegradabilityofsemanticsimilarityisubiquitous,butitismostimportantfor
understandingtheorganizationofadjectivalandadverbialmeanings.
Antonymy
Anotherfamiliarrelationisantonymy,whichturnsouttobesurprisinglydifficultto
define. Theantonymofawordxissometimesnot-x,butnotalways. Forexample,rich
andpoorareantonyms,buttosaythatsomeoneisnotrichdoesnotimplythattheymust
bepoor;manypeopleconsiderthemselvesneitherrichnorpoor. Antonymy,which
seemstobeasimplesymmetricrelation,isactuallyquitecomplex,yetspeakersof
Englishhavelittledifficultyrecognizingantonymswhentheyseethem.
Antonymyisalexicalrelationbetweenwordforms,notasemanticrelationbetween
wordmeanings. Forexample,themeanings{rise,ascend}and{fall,descend}maybe
conceptualopposites,buttheyarenotantonyms;[rise/fall]areantonymsandsoare
[ascend/descend],butmostpeoplehesitateandlookthoughtfulwhenaskedifriseand
descend,orascendandfall,areantonyms.Suchfactsmakeapparenttheneedto
distinguishbetweensemanticrelationsbetweenwordformsandsemanticrelations
betweenwordmeanings. Antonymyprovidesacentralorganizingprincipleforthe
adjectivesandadverbsinWordNet,andthecomplicationsthatarisefromthefactthat
antonymyisasemanticrelationbetweenwordsarebetterdiscussedinthatcontext.
45
-8-
Hyponymy
Unlikesynonymyandantonymy,whicharelexicalrelationsbetweenwordforms,
hyponymy/hypernymyisasemanticrelationbetweenwordmeanings:e.g.,{maple}isa
hyponymof{tree},and{tree}isahyponymof{plant}. Muchattentionhasbeen
devotedtohyponymy/hypernymy(variouslycalledsubordination/superordination,
subset/superset,orthe
ISA
relation). Aconceptrepresentedbythesynset{x,x¢,...}is
saidtobeahyponymoftheconceptrepresentedbythesynset{y,y¢,...}ifnative
speakersofEnglishacceptsentencesconstructedfromsuchframesasAnxisa(kindof)
y. Therelationcanberepresentedbyincludingin{x,x¢,...}apointertoits
superordinate,andincludingin{y,y¢,...}pointerstoitshyponyms.
Hyponymyistransitiveandasymmetrical(Lyons,1977,vol.1),and,sincethereis
normallyasinglesuperordinate,itgeneratesahierarchicalsemanticstructure,inwhicha
hyponymissaidtobebelowitssuperordinate. Suchhierarchicalrepresentationsare
widelyusedintheconstructionofinformationretrievalsystems,wheretheyarecalled
inheritancesystems(Touretzky,1986):ahyponyminheritsallthefeaturesofthemore
genericconceptandaddsatleastonefeaturethatdistinguishesitfromitssuperordinate
andfromanyotherhyponymsofthatsuperordinate. Forexample,mapleinheritsthe
featuresofitssuperordinate,tree,butisdistinguishedfromothertreesbythehardnessof
itswood,theshapeofitsleaves,theuseofitssapforsyrup,etc. Thisconvention
providesthecentralorganizingprincipleforthenounsinWordNet.
Meronymy
Synonymy,antonymy,andhyponymyarefamiliarrelations. Theyapplywidely
throughoutthelexiconandpeopledonotneedspecialtraininginlinguisticsinorderto
appreciatethem. Anotherrelationsharingtheseadvantages—asemanticrelation—isthe
part-whole(or
HASA
)relation,knowntolexicalsemanticistsasmeronymy/holonymy. A
conceptrepresentedbythesynset{x,x¢,...}isameronymofaconceptrepresentedby
thesynset{y,y¢,...}ifnativespeakersofEnglishacceptsentencesconstructedfrom
suchframesasAyhasanx(asapart)orAnxisapartofy. Themeronymicrelationis
transitive(withqualifications)andasymmetrical(Cruse,1986),andcanbeusedto
constructaparthierarchy(withsomereservations,sinceameronymcanhavemany
holonyms). Itwillbeassumedthattheconceptofapartofawholecanbeapartofa
conceptofthewhole,althoughitisrecognizedthattheimplicationsofthisassumption
deservemorediscussionthantheywillreceivehere.
Theseandothersimilarrelationsservetoorganizethementallexicon. Theycanbe
representedinWordNetbyparentheticalgroupingsorbypointers(labeledarcs)fromone
synsettoanother. Theserelationsrepresentassociationsthatformacomplexnetwork;
knowingwhereawordissituatedinthatnetworkisanimportantpartofknowingthe
word’smeaning. Itisnotprofitabletodiscusstheserelationsintheabstract,however,
becausetheyplaydifferentrolesinorganizingthelexicalknowledgeassociatedwith
differentsyntacticcategories.
22
-9-
MorphologicalRelations
Animportantclassoflexicalrelationsarethemorphologicalrelationsbetween
wordforms. Initially,interestwaslimitedtosemanticrelations;noplansweremadeto
includemorphologicalrelationsinWordNet. Asworkprogressed,however,itbecame
increasinglyobviousthatifWordNetwastobeofanypracticalusetoanyone,itwould
havetodealwithinflectionalmorphology. Forexample,ifsomeoneputthecomputer’s
cursoronthewordtreesandclickedarequestforinformation,WordNetshouldnotreply
thatthewordwasnotinthedatabase. Aprogramwasneededtostripoffthepluralsuffix
andthentolookuptree,whichcertainlyisinthedatabase. Thisneedledtothe
developmentofaprogramfordealingwithinflectionalmorphology.
AlthoughtheinflectionalmorphologyofEnglishisrelativelysimple,writinga
computerprogramtodealwithitprovedtobeamorecomplextaskthanhadbeen
expected. Verbsarethemajorproblem,ofcourse,sincetherearefourformsandmany
irregularverbs. Butthesoftwarehasbeenwrittenandispresentlyavailableaspartofthe
interfacebetweenthelexicaldatabaseandtheuser. Inthecourseofthisdevelopmentit
becameobviousthatprogramsdealingwithderivationalmorphologywouldgreatly
enhancethevalueofWordNet,butthatmoreambitiousprojecthasnotyetbeen
undertaken.
Thethreepapersfollowingthisintroductionhavelittletosayaboutlexicalrelations
resultingfrominflectionalmorphology,sincethoserelations areincorporatedinthe
interfacetoWordNet,notinthecentraldatabase.
42
-10-
NounsinWordNet:ALexicalInheritanceSystem
GeorgeA.Miller
(RevisedAugust1993)
Definitionsofcommonnounstypicallygiveasuperordinatetermplus
distinguishingfeatures;thatinformationprovidesthebasisfororganizingnoun
filesinWordNet. Thesuperordinaterelation(hyponymy)generatesa
hierarchicalsemanticorganizationthatisduplicatedinthenounfilesbythe
useoflabeledpointersbetweensetsofsynonyms(synsets). Thehierarchyis
limitedindepth,seldomexceedingmorethanadozenlevels. Distinguishing
featuresareenteredinsuchawayastocreatealexicalinheritancesystem,a
systeminwhicheachwordinheritsthedistinguishingfeaturesofallits
superordinates. Threetypesofdistinguishingfeaturesarediscussed:attributes
(modification),parts(meronymy),andfunctions(predication),butonly
meronymyispresentlyimplementedinthenounfiles. Antonymyisalsofound
betweennouns,butitisnotafundamentalorganizingprinciplefornouns.
Coverageispartitionedintotwenty-fivetopicalfiles,eachofwhichdealswith
adifferentprimitivesemanticcomponent.
Asthisiswritten,WordNetcontainsapproximately57,000nounwordforms
organizedintoapproximately48,800wordmeanings(synsets). Thenumbersare
approximatebecauseWordNetcontinuestogrow—oneadvantageofanon-linedatabase.
Manyofthesenounsarecompounds,ofcourse;afewareartificialcollocationsinvented
fortheconvenienceofcategorization. Noattempthasbeenmadetoincludeproper
nouns;ontheotherhand,sincemanycommonnounsoncewerenames,noserious
attempthasbeenmadetoexcludethem. Intermsofcoverage,WordNet’sgoalsdiffer
littlefromthoseofagoodstandardhandheldcollegiate-leveldictionary. Itisinthe
organizationofthatinformationthatWordNetaspirestoinnovation.
Ifsomeoneaskshowtouseaconventionaldictionary,itiscustomarytoexplainthe
differentkindsofinformationpackedintolexicalentries:spelling,pronunciation,
inflectedandderivativeforms,etymology,partofspeech,definitionsandillustrativeuses
ofalternativesenses,synonymsandantonyms,specialusagenotes,occasionalline
drawingsorplates—agooddictionaryisaremarkablestoreofinformation. Butif
someoneaskshowtoimproveadictionary,itbecomesnecessarytoconsiderwhatisnot
included. Andwhen,asinthecaseofWordNet,improvementsareintendedtoreflect
psycholinguisticprinciples,thefocalconcernbecomeswhatisnotincludedinthe
definitions.
Examplesofferthesimplestwaytocharacterizetheomissions. Takeonemeaning
ofthenountree,thesensehavingtodowithtreesasplants. Conventionaldictionaries
definethissenseoftreebysomesuchglossas:aplantthatislarge,woody,perennial,
andhasadistincttrunk.Ofcourse,theactualwordingisusuallymorefelicitous—a
large,woody,perennialplantwithadistincttrunk,forexample—buttheunderlying
logicisthesame:superordinateplusdistinguishers. Thepointisthattheprototypical
Documents you may be interested
Documents you may be interested