A Graph Model for Unsupervised Lexical Acquisition
Dominic Widdows and Beate Dorow
Center for the Study of Language and Information
210 Panama Street
Stanford University
Stanford CA 94305-4115
fdwiddows,beateg@csli.stanford.edu
Abstract
Thispaperpresentsanunsupervisedmethodfor
assemblingsemanticknowledgefromapart-of-
speechtaggedcorpususinggraphalgorithms.
The graph model is built by linking pairs of
wordswhichparticipateinparticularsyntactic
relationships. Wefocusonthesymmetricrela-
tionshipbetweenpairsofnounswhichoccurto-
getherinlists.Anincrementalcluster-building
algorithmusingthispartofthegraphachieves
82%accuracyatalexicalacquisitiontask,eval-
uatedagainstWordNetclasses.Themodelnat-
urallyrealisesdomainandcorpusspeci cam-
biguitiesasdistinctcomponentsinthegraph
surroundinganambiguousword.
1 Introduction
Semanticknowledgeforparticulardomainsis
increasinglyimportantinNLP.Manyapplica-
tionssuchasWord-SenseDisambiguation,In-
formationExtractionandSpeechRecognition
all require lexicons. The coverage of hand-
builtlexicalresourcessuchasWordNet(Fell-
baum,1998)hasincreaseddramaticallyinre-
cent years, but leaves several problems and
challenges. Coverage is poor in many criti-
cal,rapidlychangingdomainssuchascurrent
a airs, medicineandtechnology, wheremuch
timeisstillspentbyhumanexpertsemployed
to recognise and classify new terms. Most
languages remain poorly covered in compari-
sonwithEnglish. Hand-builtlexicalresources
whichcannotbeautomaticallyupdatedcanof-
tenbesimplymisleading. Forexample,using
WordNettorecognisethatthewordapplerefers
toafruitoratreeisagraveerrorinthemany
situationswherethiswordreferstoacomputer
manufacturer,asensewhichWordNetdoesnot
cover.ForNLPtoreachawiderclassofappli-
cationsinpractice,theabilitytoassembleand
updateappropriatesemantic knowledgeauto-
maticallywillbevital.
Thispaperdescribesamethodforarranging
semantic information into a graph (Bollob as,
1998),wherethenodesarewordsandtheedges
(also called links) represent relationships be-
tweenwords.Thepaperisarrangedasfollows.
Section 2 reviews previous work on semantic
similarityandlexicalacquisition.Section3de-
scribeshowthegraphmodelwasbuiltfromthe
PoS-taggedBritishNationalCorpus.Section4
describesanewincrementalalgorithmusedto
buildcategoriesofwordsstepbystepfromthe
graphmodel.Section5demonstratesthisalgo-
rithminactionandevaluatestheresultsagainst
WordNetclasses,obtainingstate-of-the-artre-
sults.Section6describeshowthegraphmodel
canbeusedtorecognisewhenwordsarepoly-
semousandtoobtaingroupsofwordsrepresen-
tativeofthedi erentsenses.
2 Previous Work
Mostworkonautomaticlexicalacquisitionhas
been based at some point on the notion of
semantic similarity. The underlyingclaim is
thatwordswhicharesemanticallysimilaroccur
withsimilar distributionsandinsimilarcon-
texts(MillerandCharles,1991).
Themainresultstodateinthe eldofau-
tomaticlexicalacquisitionareconcernedwith
extractinglistsofwordsreckonedtobelongto-
getherinaparticularcategory,suchasvehicles
orweapons(Rilo andShepherd,1997)(Roark
andCharniak,1998). RoarkandCharniakde-
scribea\genericalgorithm"forextractingsuch
listsofsimilarwordsusingthenotionofseman-
ticsimilarity,asfollows(RoarkandCharniak,
1998,x1).
1.Foragivencategory,chooseasmall
setofexemplars(or‘seedwords’)
2. Countco-occurrenceofwordsand
seedwordswithinacorpus
3. Usea gureofmeritbasedupon
thesecountstoselectnewseedwords
4.Returntostep2anditeratentimes
5.Usea gureofmerittorankwords
forcategorymembershipandoutputa
rankedlist
Algorithms of this type were used by Rilo 
andShepherd(1997)andRoarkandCharniak
(1998), reporting accuracies of 17% and35%
respectively. Likethealgorithmwepresentin
Section5,thesimilaritymeasure(or‘ gureof
merit’) used inthese cases was based on co-
occurrenceinlists.
Bothoftheseworksevaluatedtheirresults
byaskinghumanstojudgewhetheritemsgen-
eratedwereappropriatemembersofthecate-
goriessought. Rilo andShepherd(1997)also
givesomecreditfor‘relatedwords’(forexample
crashmightberegardedasbeingrelatedtothe
categoryvehicles).
One problem with these techniques is the
dangerof‘infections’|onceanyincorrector
out-of-category word has been admitted, the
neighboursofthiswordarealsolikelytobead-
mitted. InSection4wepresentanalgorithm
whichgoessomewaytowardsreducingsuchin-
fections.
Theearlyresultshavebeenimproveduponby
Rilo andJones(1999),wherea‘mutualboot-
strapping’approachisusedtoextractwordsin
particular semantic categories and expression
patternsforrecognisingrelationshipsbetween
thesewordsforthepurposesofinformationex-
traction. Theaccuracyachievedinthisexperi-
mentissometimesashighas78%andisthere-
forecomparabletotheresultsreportedinthis
paper.
Anotherwaytoobtainword-sensesdirectly
from corpora is to use clustering algorithms
onfeature-vectors(Lin,1998; Sch utze, 1998).
Clusteringtechniquescanalsobeusedtodis-
criminatebetweendi erentsensesofanambigu-
ousword. Ageneralproblemforsuchcluster-
ingtechniquesliesinthequestionofhowmany
clustersoneshouldhave,i.e. howmanysenses
areappropriateforaparticularwordinagiven
domain(ManningandSch utze,1999,Ch14).
Lin’sapproachtothisproblem(Lin,1998)is
tobuilda‘similaritytree’(usingwhatisinef-
fectahierarchicalclusteringmethod)ofwords
relatedtoatargetword(inthiscasetheword
duty).Di erentsensesofdutycanbediscerned
asdi erentsub-treesofthissimilaritytree.We
presentanewmethodforword-sensediscrimi-
nationinSection6.
3 Building a Graph from a
PoS-tagged Corpus
Inthissectionwedescribehowagraph|a
collection of nodesand links | was built to
representtherelationshipsbetweennouns.The
modelwasbuiltusingtheBritishNationalCor-
puswhichisautomaticallytaggedforpartsof
speech.
Initially,grammaticalrelationsbetweenpairs
ofwordswereextracted. Therelationshipsex-
tractedwerethefollowing:
 Noun(assumedtobesubject)Verb
 VerbNoun(assumedtobeobject)
 AdjectiveNoun
 NounNoun(oftenthe rstnounismodify-
ingthesecond)
 Nounand/orNoun
Thelastoftheserelationshipsoftenoccurs
whenthepairofnounsispartofalist. Since
listsareusuallycomprisedofobjectswhichare
similar in some way, these relationships have
beenusedtoextractlistsofnounswithsimilar
properties(Rilo andShepherd,1997)(Roark
andCharniak,1998). Inthispaperwetoofo-
cusonnounsco-occurringinlists. Thisisbe-
causethenounand/ornounrelationshipisthe
onlysymmetricrelationshipinourmodel,and
symmetricrelationshipsaremucheasiertoma-
nipulatethanasymmetricones.Ourfullgraph
containsmanydirectedlinksbetweenwordsof
di erent parts of speech. Initial experiments
withthismodelshowconsiderablepromisebut
areattooearlyastagetobereporteduponyet.
Thusthegraphusedinmostofthispaperrepre-
sentsonlynouns.Eachnoderepresentsanoun
andtwonodeshavealinkbetweenthemifthey
co-occurseparatedbytheconjunctionsandor
or,andeachlinkisweightedaccordingtothe
numberoftimestheco-occurrenceisobserved.
Variouscuto functionswereusedtodeter-
minehowmanytimesarelationshipmustbe
observedtobecountedasalinkinthegraph.
Awell-behavedoptionwastotakethetop n
neighboursofeachword,wherencouldbede-
termined by the user. In this way the link-
weightingschemewasreducedtoalink-ranking
scheme. Oneconsequenceofthisdecisionwas
thatlinkstomorecommonwordswerepreferred
overlinkstorarerwords. Thisdecisionmay
havee ectivelyboostedprecisionattheexpense
ofrecall,becausethepreferredlinksaretofairly
commonand(probably)morestablewords.Re-
searchisneedtorevealtheoreticallymotivated
orexperimentallyoptimaltechniquesforselect-
ingtheimportancetoassigntoeachlink|the
choicesmadeinthisareasofarareoftenofan
adhocnature.
Thegraphusedintheexperimentsdescribed
has 99,454 nodes (nouns) and 587,475 links.
There were roughly 400,000 di erent types
tagged as nouns in the corpus, so the graph
model represents about one quarter of these
nouns, including most of the more common
ones.
4 An Incremental Algorithm for
Extracting Categories of Similar
Words
Inthissectionwedescribeanewalgorithmfor
addingthe‘mostsimilarnode’toanexisting
collection ofnodesinawaywhichincremen-
tallybuildsastablecluster. Werelyentirely
uponthegraphtodeducetherelativeimpor-
tanceofrelationships. Inparticular,ouralgo-
rithmisdesignedtoreduceso-called‘infections’
(RoarkandCharniak,1998,x3)wheretheinclu-
sionofanout-of-categorywordwhichhappens
toco-occurwithoneofthecategorywordscan
signi cantlydistortthe nallist.
Hereistheprocessweusetoselectandadd
the‘mostsimilarnode’toasetofnodes:
De nition1Let A be a set of nodes and
let N(A), the neighbours of A, be the nodes
whicharelinkedtoany a 2 A. (So N(A)=S
a2AN(a).)
Thebestnewnodeistakentobethenode
b2N(A)nAwiththehighestproportionoflinks
toN(A).Moreprecisely,foreachu2N(A)nA,
letthea nitybetweenuandAbegivenbythe
ratio
jN(u)\N(A)j
jN(u)j :
Thebestnewnode b 2 N(A)nA isthenode
whichmaximisesthisa nityscore.
Thisalgorithmhasbeenbuiltintoanon-line
demonstration where the user inputs a given
seedwordandcanthenseetheclusterofre-
latedwordsbeinggraduallyassembled.
The algorithm is particularly e ective at
avoiding infections arising from spurious co-
occurrencesandfromambiguity. Consider,for
example,thegraphbuiltaroundthewordap-
pleinFigure6.Supposethatwestartwiththe
seed-listapple,orange,banana. Howevermany
timesthestring\AppleandNovell"occursin
thecorpus,thenovellnodewillnotbeadded
tothislistbecauseitdoesn’thavealinktoor-
ange,bananaoranyoftheirneighboursexcept
forapple. Onewaytosummarisethee ectof
thisdecisionisthatthealgorithmaddswords
toclustersdependingontypefrequencyrather
thantokenfrequency.Thisavoidsspuriouslinks
dueto(forexample)particularidiomsrather
thangeniunesemanticsimilarity.
5 Examples and Evaluation
Inthissectionwegiveexamplesoflexicalcat-
egoriesextractedbyourmethodandevaluate
themagainstthecorrespondingclassesinWord-
Net.
5.1 Methodology
Our methodology is as follows. Consider an
intuitive category of objects such as musical
instruments. De ne the ‘WordNet class’ or
‘WordNetcategory’ofmusicalinstrumentsto
bethecollectionofsynsetssubsumedinWord-
Netbythemusicalinstrumentssynset. Takea
‘protypical example’ of a musical instrument,
suchas piano. Thealgorithm de nedin (1)
givesawayof ndingthennodesdeemedtobe
mostcloselyrelatedtothepianonode. These
canthenbechecked toseeiftheyaremem-
bers of the WordNet class of musical instru-
ments.Thismethodiseasiertoimplementand
lessopentovariationthanhumanjudgements.
WhileWordNetoranyotherlexicalresourceis
notaperfectarbiter,itishopedthatthisexper-
imentprocedureisbothreliableandrepeatable.
Thetenclassesofwordschosenwerecrimes,
places, tools, vehicles, musical instruments,
clothes,diseases,bodyparts,academicsubjects
andfoodstu s. Theclasseswerechosenbefore
theexperimentwascarriedoutsothatthere-
sultscouldnotbemassagedtoonlyusethose
classeswhichgavegoodresults.(The rst4cat-
egoriesarealsousedby(Rilo andShepherd,
1997)and(RoarkandCharniak,1998)andso
wereincludedforcomparison.) Havingchosen
theseclasses,20wordswereretrievedusinga
singleseed-wordchosenfromtheclassinques-
tion.
Thislistofwordsclearlydependsontheseed
wordchosen. Whilewehavetriedtooptimise
thischoice,itdependsonthecorpusandthe
themodel. Thein uenceofsemantic Proto-
typeTheory(Rosch,1988)isapparentinthis
process,alinkwewouldliketoinvestigatein
moredetail.Itispossibletochooseanoptimal
seedwordforaparticularcategory:itshouldbe
possibletocomparetheseoptimalseedwords
withthe‘prototypes’suggestedbypsychologi-
calexperiments(MervisandRosch,1981).
5.2 Results
Theresultsforalistoftenclassesandproto-
typicalwordsaregiveninTable1.Wordswhich
arecorrectmembersoftheclassessoughtare
in Roman type: incorrect results are in ital-
ics. Thedecisionbetweencorrectnessandin-
correctnesswasmadeonastrictbasisforthe
sakeofobjectivityandtoenabletherepeata-
bility ofthe experiment: wordswhichare in
WordNetwerecountedascorrectresultsonlyif
theyareactualmembersoftheWordNetclass
inquestion. Thusbrigandageisnotregarded
asacrimeeventhoughitisclearlyanactof
wrongdoing,orchestraisnotregardedasamu-
sicalinstrumentbecauseitisacollectionofin-
strumentsratherthanasingleinstrument,etc.
Theonlyexceptionswehavemadearetheterms
wyndandplanetology(markedinbold),which
arenotinWordNetbutarecorrect nonethe-
less. Theseconditionsareatleastasstringent
asthoseofpreviousexperiments,particularly
thoseofRilo andShepherd(1997)whoalso
givecreditforwordsassociatedwithbutnot
belongingtoaparticularcategory.(Ithasbeen
pointedoutthatmanypolysemouswordsmay
occurinseveralclasses,makingthetaskeasier
becauseformanywordsthereareseveralclasses
whichouralgorithmwouldgivecreditfor.)
With these conditions, our algorithm re-
trievesonly36incorrecttermsoutofatotal
of200,givinganaccuracyof82%.
5.3 Analysis
Ourresultsareanorderofmagnitudebetter
than those reported by Rilo and Shepherd
(1997) and Roark andCharniak (1998), who
reportaverageaccuraciesof17%and35%re-
spectively. (Ourresultsarealsoslightlybetter
thanthosereportedbyRilo andJones(1999)).
Since the algorithms used are in many ways
verysimilar,thisimprovementdemandsexpla-
nation.
Someofthedi erenceinaccuracycanbeat-
tributedtothecorporaused.Theexperiments
in(Rilo andShepherd,1997)wereperformed
onthe500,000wordMUC-4corpus,andthose
of(RoarkandCharniak,1998)wereperformed
usingMUC-4andtheWallStreetJournalcor-
pus(some30millionwords). Ourmodelwas
built using the British National Corpus (100
millionwords). Ontheotherhand,ourmodel
wasbuiltusingonlyapart-of-speechtaggedcor-
pus.Thehighaccuracyachievedthusquestions
theconclusiondrawnbyRoarkandCharniak
(1998)that‘parsingisinvaluable’. Ourresults
clearlyindicatethatalargePoS-taggedcorpus
maybemuchbetterforautomaticlexicalac-
quisitionthanasmallfully-parsedcorpus.This
claimcouldofcoursebetestedbycomparing
techniquesonthesamecorpus.
ToevaluatetheadvantageofusingPoSinfor-
mation,wecomparedthegraphmodelwitha
similaritythesaurusgeneratedusingLatentSe-
manticIndexing(ManningandSch utze,1999,
Ch15),a‘bag-of-words’approach,onthesame
corpus. The same number of nouns was re-
trieved for each class using the graph model
andLSI.TheLSIsimilaritythesaurusobtained
anaccuracyof31%,muchlessthanthegraph
model’s 82%. This is because LSI retrieves
wordswhicharerelatedbycontextbutarenot
inthesameclass: forexample,theneighbours
ofpianofoundusingLSIcosine-similarityonthe
BNCcorpusincludewordssuchascomposer,
music,Bach,concertoanddance,whicharere-
lated butcertainly not inthe same semantic
class.
TheincrementalclusteringalgorithmofDef-
inition(1)workswellatpreventing‘infections’
Class SeedWord NeighboursProducedbyGraphModel
crimes murder crimetheftarsonimportuningincestfraudlarcenyparricide
burglaryvandalismindecencyviolenceo encesabusebrig-
andagemanslaughterpillageraperobberyassaultlewdness
places park pathvillagelaneview eldchurchsquareroadavenuegarden
castlewyndgaragehousechapeldrivecrescenthomeplace
cathedralstreet
tools screwdriver chiselnavillenailshoulderknifedrillmatchstickmorgenthau
gizmohandkneeelbowmalletpenknifegallielegarmsickle
bolsterhammer
vehicle
conveyance
train tramcardriverpassengerscoachlorrytruckaeroplanecoons
planetrailerboattaxipedestriansvansvehiclesjeepbusbuses
helicopter
musical
instruments
piano fortepianoorchestramarimbaclarsachviolincizekviolaoboe
 utehornbassoonculbonemandolinclarinetequiluzcontra-
basssaxophoneguitarcello
clothes shirt chapeaubrascardigantrousersbreechesskirtjeansbootspair
shoesblousedresshatwaistcoatjumpersweatercoatcravat
tieleggings
diseases typhoid malariaaidspoliocancerdiseaseatelectasisillnessescholera
hivdeathsdiphtheriainfectionshepatitistuberculosiscirrho-
sisdiptheriabronchitispneumoniameaslesdysentery
bodyparts stomach headhipsthighsneckshoulderschestbackeyestoesbreasts
kneesfeetfacebellybuttockshawsankleswaistlegs
academic
subjects
physics astrophysicsphilosophyhumanitiesartreligionsciencepol-
itics astronomy sociology chemistry history theology eco-
nomicsliteraturemathsanthropologyculturemathematics
geographyplanetology
foodstu s cake macaroonsconfectioneriescreamrollssandwichescroissant
bunssconescheesebiscuitdrinkspastriesteadanishbutter
lemonadebreadchocolateco eemilk
Table1:Classesofsimilarwordsgivenbythegraphmodel.
andkeepingclusterswithinoneparticularclass.
Thenotableexceptionisthetoolsclass,where
thewordhandappearstointroduceinfection.
Inconclusion,itisclearthatthegraphmodel
combinedwiththeincrementalclusteringalgo-
rithmofDe nition1performsbetterthanmost
previousmethodsatthetaskofautomaticlex-
icalacquisition.
6 Recognising Polysemy
Sofarwehavepresentedagraphmodelbuilt
uponnounco-occurrencewhichperformsmuch
betterthanpreviouslyreportedmethodsatthe
taskofautomaticlexical acquisition. Thisis
animportanttask,becauseassemblingandtun-
inglexiconsforspeci cNLPsystemsisincreas-
ingly necessary. We nowtake astepfurther
andpresentasimplemethodfornotonlyas-
semblingwordswithsimilarmeanings,butfor
empiricallyrecognisingwhenawordhasseveral
meanings.
Recognising and resolving ambiguity is
an important task in semantic processing.
The traditional Word Sense Disambiguation
(WSD)problemaddressesonlytheambiguity-
resolutionpartoftheproblem:compilingasuit-
ablelistofpolysemouswordsandtheirpossible
sensesisataskforwhichhumansaretradition-
allyneeded(Kilgarri andRosenzweig,2000).
ThismakestraditionalWSDanintensivelysu-
pervisedandcostlyprocess. Breadthofcover-
agedoesnotinitselfsolvethisproblem:general
lexicalresourcessuchasWordNetcanprovide
toomanysensesmanyofwhicharerarelyused
inparticulardomainsorcorpora(Galeetal.,
1992).
Thegraphmodelpresentedinthispapersug-
gestsanewmethodforrecognisingrelevantpol-
ysemy. Wewillneedasmallamountoftermi-
nologyfromgraphtheory(Bollob as,1998).
De nition2(Bollob as, 1998, Ch 1 x1)
LetG=(V;E)beagraph,whereV istheset
ofvertices(nodes)ofGandE V  V isthe
setofedgesofG.
 Twonodesv1;vn aresaidtobeconnected
ifthereexistsapathfv1;v2;:::;vn 1;vng
suchthat(vj;vj+1)2Efor1 j<n.
 Connectednessisanequivalencerelation.
 TheequivalenceclassesofthegraphGun-
derthisrelationarecalledthecomponents
ofG.
Wearenowinapositiontode nethesenses
ofawordasrepresentedbyaparticulargraph.
De nition3LetGbeagraphofwordsclosely
relatedtoaseed-word w,andletGnw bethe
subgraphwhichresultsfromtheremovalofthe
seed-nodew.
The connected components of the subgraph
Gnwarethesensesofthewordwwithrespect
tothegraphG.
Asanillustrativeexample,considerthelocal
graphgeneratedforthewordapple(6).There-
movaloftheapplenoderesultsinthreeseparate
componentswhichrepresentthedi erentsenses
ofapple:fruit,trees,andcomputers.De nition
3givesanextremelygoodmodelofthesenses
ofapplefoundintheBNC.(Inthiscasebetter
thanWordNetwhichdoesnotcontainthevery
commoncorporatemeaning.)
Theintuitivenotionofambiguitybeingpre-
sentedisasfollows. Anambiguouswordoften
connectsotherwiseunrelatedareasofmeaning.
De nition3recognisestheambiguityofapple
becausethiswordislinkedtobothbananaand
novell,wordswhichotherwisehavenothingto
dowithoneanother.
It is well-known that any graph can be
thoughtofasacollectionoffeature-vectors,for
examplebytakingtherow-vectorsintheadja-
cencymatrix(Bollob as,1998,Ch2x3). There
mightthereforebefundamentalsimilaritiesbe-
tweenourapproachandmethodswhichrelyon
similaritiesbetweenfeature-vectors.
Extra motivation for thistechnique is pro-
vided by Word-Sense Disambiguation. The
standardmethodforthistaskistousehand-
labelled data to train a learning algorithm,
whichwilloftenpickoutparticularwordsas
Bayesianclassi erswhichindicateonesenseor
theother. (Soifmicrosoftoccursinthesame
sentenceasapplewemighttakethisasevidence
thatappleisbeingusedinthecorporatesense.)
Clearly,thewordsinthedi erentcomponents
inDiagram6canpotentiallybeusedasclassi-
 ersforjustthispurpose,obviatingtheneedfor
time-consuminghumanannotation. Thistech-
niquewillbeassessedandevaluatedinfuture
experiments.
Demonstration
Anonlineversionofthegraphmodelandthein-
crementalclusteringalgorithmdescribedinthis
paperarepubliclyavailable1fordemonstration
purposesandtoallowuserstoobservethegen-
eralityofourtechniques. Asampleoutputis
includedinFigure6.
Acknowledgements
Theauthorswouldliketothanktheanonymous
reviewerswhosecommentswereagreathelpin
makingthispapermorefocussed: anyshort-
comingsremainentirelyourownresponsibility.
Thisresearchwassupportedinpartbythe
ResearchCollaborationbetweentheNTTCom-
municationScienceLaboratories,NipponTele-
graph and Telephone Corporation and CSLI,
StanfordUniversity,andbyEC/NSFgrantIST-
1999-11438fortheMUCHMOREproject. 2
1http://infomap.stanford.edu/graphs
2http://muchmore.dfki.de
Figure1:Automaticallygeneratedgraphshow-
ing the word apple and semantically related
nouns

References

B ela Bollob as. 1998. Modern Graph Theory.
Number184inGraduatetextsinMathemat-
ics.Springer-Verlag.

ChristianeFellbaum.1998.WordNet:Anelec-
tronic lexical database. MIT press, Cam-
bridgeMA.

W.Gale,K.Church,andD.Yarowsky. 1992.
Onesenseperdiscourse. InDARPAspeech
andNaturalLanguageWorkshop,Harriman,
NY.

AdamKilgarri andJosephRosenzweig.2000.
English senseval: report and results. In
LREC,Athens.

Dekang Lin. 1998. Automatic retrieval and
clustering of similar words. In COLING-
ACL,Montreal,August.

ChristopherD.ManningandHinrichSch utze.
1999.FoundationsofStatisticalNaturalLan-
guage Processing. The MIT Press, Cam-
bridge,Massachusetts.

C.MervisandE.Rosch. 1981. Categorization
ofnaturalobjects.AnnualReviewofPsychol-
ogy,32:89{115.

GeorgeA.MillerandWilliamG.Charles.1991.
Contextualcorrelatesofsemanticsimilarity.
LanguageandCognitiveProcesses,6(1):1{28.

Ellen Rilo and Rosie Jones. 1999. Learn-
ingdictionariesforinfomationextractionby
multi-levelbootstrapping. InProceedingsof
theSixteenthNationalConferenceonArti -
cialIntelligence,pages472{479.AAAI.

Ellen Rilo and Jessica Shepherd. 1997. A
corpus-based approach for building seman-
tic lexicons. In Claire Cardie and Ralph
Weischedel,editors,ProceedingsoftheSecond
ConferenceonEmpiricalMethodsinNatural
LanguageProcessing,pages117{124.Associ-
ationforComputationalLinguistics,Somer-
set,NewJersey.

Brian Roark and Eugene Charniak. 1998.
Noun-phraseco-occurencestatisticsforsemi-
automaticsemanticlexiconconstruction. In
COLING-ACL,pages1110{1116.

E.Rosch.1988.Principlesofcategorization.In
A.CollinsandE.E.Smith, editors, Read-
ingsinCognitiveScience:APerspectivefrom
PsychologyandArti cialIntelligence,pages
312{322.Kaufmann,SanMateo,CA.

HinrichSch utze. 1998. Automaticwordsense
discrimination. Computational Linguistics,
24(1):97{124.
