面向微博数据的命名实体识别研究与实现

《面向微博数据的命名实体识别研究与实现》由会员上传分享，免费在线阅读，更多相关内容在教育资源-天天文库。

1、万方数据东北大学壁士堂堡垒圭一!堕——●_-—__-__———●__●●-——_-_—___—●__—●_-—__-—___—●_——_●h—●_———●-—-●——__-__--—___●_-—--—_-_———————一一一万方数据onofMicroblogDataAbstractWithpopularityofmicroblog，microbloghasbecomeanewsocialmesaofinformationreleaseandpropagationg．BytheendofDecember2012，Sinamicrob

2、logreachedmorethan500millionregisteredusers．Microblogdataincreasedalsoinwhichthereismuchvaluableinformationfororganizationandindividual．Thus，informationextraction，analysisandnaturallanguageprocessingcarriedoutmicroblogbecomearesearchhotspot．Namedentityrecognitionisparti

3、cularlyimportantasabasicmissionoftheseresearches，butcurrentnamedentityrecognitionmicroblogresearchisnotverymature．Tranditionalmethodsofnamedentityrecognitionmicroblogdatacannotobtainsatisfactoryresult，whichhindersthefollow-upwork．Thispapermainlyresearchesnamedentityreco

4、gnitionmicroblogdata．Characteristicsofmicroblogdataleadtothefailureoftraditionalmethods．Thefundamentalreasonsarelistedinfourpoints．First，eachmicroblogisveryshortwithlimitedinformation，whichmakesitdifficulttofullyintegratealotofrelevantinformation．Second，thereismuchnoisy

5、data，butnoisyimmunityofmodelsislow．Thesemakeoverfittingphenomenonhappenmodelsintrainingcourse．Third，thereisnosufficientcompletetrainingresourceinmicroblogdata，whichmayleadtounder-trainingofmodels，Furthermore，italwaysneedsmuchmanualwork．Fourth，fastinformationupdateofmicr

6、oblogdatamakesunderfittingphenomenonhappenintrainingcourselowadaptabilityofmodels．ExperimentshowsthatF1measureofnamedentityrecognitionresultsbyconventionalmethodswillalmostdrop20percents．Inordertosolveproblemslistedabove，thepaperimplementsnamedentityrecognitionmicroblog

7、datawithmanytechnologies．Theprecision，recallandF1measureoftheresearchresultmicroblogbackgroundgets83．7％．79．8％and81．8％respectively．Theresultimprovesmuchrelativetoconventionalmethods．Thepaperovercomesdisadvantagesinconventionalmethodsbyfollowingaspects．First，theresearchbu

8、ildsasemi·supervisednamedentityrecognitionframe，whichfiguresoutthelackoftrainingdatabyrepeating万方数据东北大学硕士学位论文A