everydogprobablychasedsomewhitecatFullnotation:

TOPh1

LZT

prpstn

every

dogprobably

chaserelHNLh12

EVENTe2TENSEpast

MOODindic

ARG1x4ARG2x13

HNLh14BVx13RESTRh15BODYh16

HNLh18ARGx13

HNLh18INSTx13

H-CONS

h5qeqh8h10qeqh12h15qeqh18h21qeqh9

Unscopedform,abbreviatednotation:

prpstneverysomewhite

dog

cat

probably

chase

Scopedforms:

somewhitecatchaseprpstnprobablyeverydog

prpstneverydogprobablysomewhitecatchaseprpstneverydogsomewhitecatprobablychaseprpstnprobablysomewhitecateverydogchaseprpstnsomewhitecatprobablyeverydogchaseprpstnsomewhitecateverydogprobablychase

Figure1:ExampleofMRSrepresentationproducedbytheERG

‘wrong’choice(i.e.,onenotlicensedbythegrammar).Onedangerwiththisisthattheinputmightbetoounderspeci- ed,buttheapproachwearetakingtothisistoletcorpusdataactasanoracletoguideagenerator.Speci cally,inthecontextofthespeechprosthesisprojectmentionedin2.3.,theinputtothegeneratormayunderspecifyclosed-classwords,suchasdeterminers.Althoughthegrammarprovidessomeconstraints(e.g.muchmaynotoccurwithpluralnouns),atotallyunderspeci eddeterminerwouldgenerallyresultingenerationoffarmorestringsthantheapplicationcouldplausiblyrequire.Frequencyinformationfromcorporacanbeusedtomakeabestguessinanap-plicationsuchasthis,wherecompleteprecisionisnotarequirement.

Weneedtodomoreworktode neappropriategeneral-purposeinterfacestothesemantics,especiallyforapplica-tionsthatusethegrammarforgeneration.However,webelieveMRSwilleventuallybehighlysuitableasarepre-sentationforasharedresourcegrammar.

4.

Ef ciencyimprovementsviashared

resources

TheLinGOERGhasbeencentralinacollaborativeef-forttoimprovetechnologyforprocessingHPSG(andsim-ilarformalisms).Untilrecently,ithasbeenarguedthathigh-levelgrammarformalismsaretooslowforreal-worldapplications.GroupsfromtheDFKI,SaarlandUniversity,theUniversityofTokyoandCSLIhavecollaboratedonim-provingef ciencybyusingtheLinGOgrammarasacom-monreferencepoint.Allthesitesusedthe[incrtsdb()]systemtomeasurevariousparametersinparsingwiththeERGoncommontest-suites.TheLKBsystemwasusedasabaselinetovalidatecorrectness.Itwasalsousedtopreprocessthetypehierarchyandlexicontoavoidthene-cessityforeachgrouptowriteaparserforthesyntaxusedintheERG.Practicallyspeaking,thisismuchfasterthanattemptingtogetgroupstoagreeonacommonsyntaxfor

thegrammarde nition les.Similarly,theothersystemshavemadeuseofvariousexpandedformsofthegrammarthatcanbeoutputbytheLKB,enablingthemtobypassprocessingstageswhichareirrelevantforcoreparsercom-parison,suchasexpansionofthetypehierarchytoformasemi-latticeandmorphologicalprocessing.

SomesmallchangestotheERGweremadetoallowthiscomparison:forinstance,therewereafewplaceswheretheLKB’sassumptionthatfeaturestructuresmaynotcontaincycleswasusedinordertoruleoutotherwisevaliduni -cations.However,itwasnotdif culttoremovethesecasesandtheERGitselfisnowneutralwithrespecttocyclicityassumptionsinthatitdoesnotgeneratecyclicstructures(modulobugs).

ThecombinationoftheLinGOERGandthe[incrtsdb()]test-suitemachineryhasenabledmuchmorede-tailedcross-platformperformanceevaluationonrealisticgrammarsthanhaspreviouslybeenpossible.Theresulthasbeenacombiningofprocessingtechniquesdevelopedbythevariousgroupstomutualbene t.(Kieferetal.,1999)isapartialreportonsomeofthiswork;(Flickingeretal.,2000)containsdetaileddiscussionsbymostoftheparticipantsinthiscollaboration.TheLKBnowincorpo-ratestechniquesthatwereadaptedfromothersystems,es-peciallyPAGEandPET(Callmeier,2000).

5.Othercollaborationsandfuturework

Inadditiontotheworkonprocessingef ciencyandper-formancepro lingdescribedabove,theERGandtheLKBalsoformthebasisofwell-establishedcollaborationsonex-tractionofstochasticlexicalizedtreegrammars(Neumann,1997)andtheintegrationwithdiscoursemodelsandprag-matics(CopestakeandLascarides,1998).EmergingjointprojectsincludetheadaptationoftheERGlexicaltypehi-erarchytolargeEnglishandJapaneselexiconsdevelopedforMTincollaborationwithFrancisBondandotherre-searchersfromNTT.Weareactivelyinvestigatingpossible