1999).CurrentlythemainLinGOERGdoesnotusede-faultsbutMaloufhasdevelopedaversionthatdoes(Mal-ouf,1999),whichdemonstratesthatitispossibletocon-siderablysimplifythetypehierarchy.Thereisatrade-offsincetheuseofdefaultsmeanstheformalismismorecom-plexandlesslikeHPSGasitisstandardlypresented.How-ever,webelievethatoveralldefaultsmakethegrammarmucheasiertounderstand.Infact,extensiveuseismadeofthedefaultformalisminrecenttheoreticalworkonHPSG(GinzburgandSag,toappear).

Despitethemanyaidstogrammarengineeringthathavebeendeveloped,wethinkthattosomeextentitjusthastobeacceptedthatitreallyisinherentlydif cult.Gram-mardevelopersareattemptingtoconstructamodelwhichmirrors(someaspectsof)areal-worldphenomenonwhichisnotatallwellunderstood.Thusgrammardevelopmentismuchmoreofaresearchenterprisethanconventionalsoftwareengineeringis.Itnecessarilyinvolvesfrequentredesignandrecodingoflargecomponentsofthegram-mar(withintheboundariesimposedbytheinitiallinguisticframework).

Asa rststeptobeingabletoworkontheLinGOERG,anypotentialgrammardeveloperneedstohaveagoodun-derstandingoflinguisticsingeneralandHPSGinparticu-lar.Thisimposesahighinitialbarrier,especiallybecausetheHPSGframeworkchangesrapidly:knowledgeofthestandardreferences(PollardandSag,1987;PollardandSag,1994)isonlyoflimiteduse.Furthermore,eventhoughtheformalismimplementedintheLKBsystemisrelativelyclosetotheformalismassumedinmosttheoreticalworkonHPSG,thereisstillaconsiderableelementofdesignnec-essaryinordertoactuallygetananalysistowork.How-ever,themosttime-consumingpartofgrammardevelop-mentisnotimplementingthewell-workedoutphenom-enadescribedinexistingworkonHPSG,butextendingtheanalysistodealwiththeconstructionswhichareneededtohandlerealtextorspeech.Thisinvolvesconsiderablelin-guisticresearch,beforeanyattemptatcodingcanbemade.Itisalsonecessarytoensurethattheresultcanbeintegratedwiththerestoftheframework,whichmayinvolvenon-trivialmodi cationtothecorecomponent.Thepositivesideofthisisthatpeoplewhohavesuf cientunderstand-ingoflinguisticstobegoodgrammardevelopers ndtheresearchcomponentinterestingandmotivating.Similarly,thisisagoodreasontodevelopagrammarusingaframe-workthatisrelativelypopularamonglinguists.Infact,totheextentthattheERGisanembodimentofresearchonHPSG,itisadesirablegoalforstudentstogaindetailedunderstandingofasubstantialportionofit,ratherthanbe-inglimitedtoasmallmodule.SeverallinguisticspaperswhichhaveresultedfromworkontheLinGOERG(Ben-derandFlickinger,1999;BenderandFlickinger,inpress;Smith,1999).

3.3.MRSrepresentation

TheMRSrepresentationisusedintheLinGOERGand,inasimpli edform,inseveralteachinggrammars.5MRSisdiscussedindetailin(Copestakeetal.,1999).Itsmost

MRS(UMRS)wasdevelopedforaGermangrammaratIBMHei-delberg.UMRSisdescribedin(EggandLebeth,1995;Egg,1998).(Abbetal.,1996)discusstheuseofUMRSinsemantictransfer(alsosee(Copestakeetal.,1995)forsemantictransferusingMRS).6

Weactuallyrelaxthisslightlyinordertotreatfragmentsofsentencessincewecurrentlyregardanatomicabbreviationofthesyntaxofthephrase,suchasNP,PPetc,aspartoftheinterface.Inthefuture,wewillalsoneedtoconsiderdiscourseinformationandinformationstructuremorecarefully.