An Overview of Probabilistic Tree Transducers for Natural Language Processing英文文献资料

《An Overview of Probabilistic Tree Transducers for Natural Language Processing英文文献资料》由会员上传分享，免费在线阅读，更多相关内容在学术论文-天天文库。

1、AnOverviewofProbabilisticTreeTransducersforNaturalLanguageProcessingKevinKnightandJonathanGraehlInformationSciencesInstitute(ISI)andComputerScienceDepartment,UniversityofSouthernCalifornia{knight,graehl}@isi.eduAbstract.Probabilisticfinite-statestringtransducers(FSTs)areextremelypop-ularinnatu

2、rallanguageprocessing,duetopowerfulgenericmethodsforap-plying,composing,andlearningthem.Unfortunately,FSTsarenotagoodfitformuchofthecurrentworkonprobabilisticmodelingformachinetranslation,summarization,paraphrasing,andlanguagemodeling.Thesemethodsoperatedi-rectlyontrees,ratherthanstrings.Wesho

3、wthattreeacceptorsandtreetransduc-erssubsumemostofthiswork,andwediscussalgorithmsforrealizingthesamebenefitsfoundinprobabilisticstringtransduction.1StringsManynaturallanguageproblemshavebeensuccessfullyattackedwithfinite-statema-chines.Ithasbeenpossibletobreakdownverycomplexproblems,bothconcept

4、uallyandliterally,intocascadesofsimplerprobabilisticfinite-statetransducers(FSTs).Thesetransducersarebidirectional,andtheycanbetrainedonsampleinput/outputstringdata.Byaddingaprobabilisticfinite-stateacceptor(FSAs)languagemodeltooneendofthecascade,wecanimplementprobabilisticnoisy-channelmodels.1

5、Fig-ure1showsacascadeofFSAsandFSTsfortheproblemoftransliteratingnamesandtechnicaltermsacrosslanguageswithdifferentsoundsandwritingsystems[1].Thefinite-stateframeworkispopularbecauseitofferspowerful,genericoperationsforstatisticalreasoningandlearning.Therearestandardalgorithmsfor:–intersectiono

6、fFSAs–forwardapplicationofstringsandFSAsthroughFSTs–backwardapplicationofstringsandFSAsthroughFSTs–compositionofFSTs–k-bestpathextraction–supervisedandunsupervisedtrainingofFSTtransitionprobabilitiesfromdata1Inthenoisy-channelframework,welookfortheoutputstringthatmaximizesP(output

7、input),whic

8、hisequivalent(byBayesRule)tomaximizingP(output)·P(input

9、output).ThefirsttermoftheproductisoftencapturedbyaprobabilisticFSA,thesecondtermbyaprobabilisticFST(oracascadeofthem).A.Gelbukh(Ed.):CICLing2005,LNCS3406,pp.1–24,2005.
cSpringer-VerlagBer