Effective Exploitation of Parallelism in NLP

《Effective Exploitation of Parallelism in NLP》由会员上传分享，免费在线阅读，更多相关内容在行业资料-天天文库。

1、EffectiveExploitationofParallelisminNLPM.P.vanLohuizenDept.ofInformationTechnologyandSystemsDelftUniversityofTechnologyZuidplantsoen4,2628BZDelft,TheNetherlandsmpvl@acm.orgKeywords:NaturalLanguageProcessing,ParallelParsing,Fine-GrainedParallelismAbstractseenasanextrac

2、onstraintonthecontext-freeback-bone.Often,theunificationoperations(whichareTheincreasingdemandforaccuracyandrobust-similartothosefoundinProlog)arethemostcostlynessfortoday'snaturallanguageprocessingsystemspartofparsing.Basically,onecantaketwoap-bringsonanincreasingdema

3、ndforcomputingpower.proachestoparallelizingaunification-basedparser.Experimentsonourunification-basedparserindicatedFirst,onecantakeunification(andsubsumption)asthatonlyfine-grainedparallelismcouldyieldaccept-anatomicoperation,andexploittheparallelismcon-ablespeedup.Wepre

4、sentaparallelizationtechniquetainedinthecontext-freepartofparsing.Second,onetoexploitthisfine-grainedparallelism.Althoughthecouldparallelizetheunificationoperationitself.Mostapproachwasdesignedfortheparallelizationofaresearchhasfocussedontheformerapproach.unification-bas

5、edparser,mostofthecomplexityiscapturedinanindependentmodulethatcanbeusedforotherapplicationsaswell.OneofthereasonsforthischoiceisthatunificationisnotinNick'sClassandthereforegenerallyconsid-1Introductioneredtobeahardtoparallelizeproblem[1,2].In[3],however,itisshownthat

6、thispropertydoesnotmeanTheperformanceofnaturallanguageprocessingitisnotfeasibletomakeanefficientparallelunifica-(NLP)applicationshasalwaysbeenlimitedbythetionalgorithm(seealsoremarksin[4]).Thereare,speedoftheworkstationsathand.Wheneverfasterhowever,otherreasonthatmaketh

7、eformerapproachcomputersenteredthemarket,NLPapplicationsweremoreattractive.Theparallelisminherentintheuni-extendedtodealwithevenmorecomplexlinguisticficationalgorithmisveryfine-grained.Thisensuresphenomenaandanevenlargerdomainofdiscourse.thatworkcanbebalancedevenlyamong

8、stprocessors,Withtheintroductionofrobustparsingtechniquesbutalsoincreasesthepotentialoverheadofthesys-(parsingofpartiallycor