资源描述：
《The Present Use of Statistics in the Evaluation of NLP Parsers》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、
2、ThePresentUseofStatisticsintheEvaluationofNLPParsersIiJ.EntwisleD.M.W.PowersIFlindersUniversityofSouthAustraliaFlindersUniversityofSouthAustraliajim@cs,flinders,edu.aupowers~cs,flinders,edu.aumAbstracthybrids]seemcapableofexceedingthe97%barrier....Next,anewsystem
3、..usesWeareconcernedthatthequalityofresultsonlylinguisticdistributionalrules.TestedproducedbyanNLPparserbearslittle,ifagainsta38,000-wordcorpusofpreviouslyany,relationtothepercentage-resultsclaimedbythevariousNLPparser-systemspresentlyunseentext,thetaggerreachesab
4、etterac-mavailableforuse.Toillustratethisproblem,curacythanprevioussystems(over99%).weexamineonereadilyavailableNLPtagging(Voutilainen,1995)andparsingsystem,theENGCGparser;andmonetagger,theBrintagger.WenoteresponsesWeareconcernedaboutthemisleadingnatureoftobothart
5、ificiallygeneratedandnaturallyoc-suchpublishedstatistics,althoughresearcherswork-curringtext.ThepercentageassessmentsareingonNLPsystemsareofcoursewellawarethatmethodologicallyflawed,andshouldbetakenthefiguresmustbeinterpretedcarefully.Tobeablemwithagrainofsalt;ins
6、tead,assessmentoftheperformanceofanNLPparsershouldbeef-toshowourcauseforconcernwemustlookatthefectedbyauser,andsolelyfromaconsidera-statisticsofawellknownparser,andalsoatitsac-tionoftheresultingparsesofexactlythein-tions.Unfortunatelytherearerelativelyfewsystemsmp
7、utwhichanNLPuserdecidestocontributewhicharefreelyavailableforconsultationorexami-forsuchanassessment.Carefulattentiontonation,andforthisreasonweareforcedtopickoninputofwhatevercorpustheuserdecideson,ispresentlytheonlysuitablequalifyingtestofsomeofthefewsystemsthat
8、are.Thecriticismsweparsingability.Theparsersavailablearenonemakeherearenotdirectedagainsttheparsersweofthemperfectibleyet,despiteapparentyieldsuseasexamples,butagainstthewayinwhichournowquotedat99%+.Weconsidertheimpactfieldistreatingitsstatistics.ofZipf'sargumento
9、f'leasteffort'onpercent-OurmainexampleistheENGCGparser,inoneageassessment;andweopenadiscussiononestimatingtherelativecomplexitiesofcorpora.specificformw
