Page 70 - 2023年第54卷第7期
P. 70
OL].(2022 - 04 - 21)[2022 - 06 - 17].https:??arxiv.org?abs?2204.10019.
[26] KENTONJDM W C,TOUTANOVALK.Bert:Pre - trainingofdeepbidirectionaltransformersforlanguageun
derstanding [C]??ProceedingsofnaacL - HLT.2019.
[27] WEIJ,ZOUK.Eda:Easydataaugmentationtechniquesforboostingperformanceontextclassificationtasks[EB?
OL].(2019 - 08 - 25)[2022 - 09 - 10].https:??arxiv.org?abs?1901.11196.
[28] KARIMIA,ROSSIL,PRATIA.Aeda:Aneasierdataaugmentationtechniquefortextclassification[EB?OL].
(2021 - 08 - 30)[2022 - 09 - 10].https:??arxiv.org?abs?2108.13230.
[29] LIUK,LIUX,YANGA,etal.Arobustadversarialtrainingapproachtomachinereadingcomprehension[C]??
ProceedingsoftheAAAIConferenceonArtificialIntelligence.2020.
[30] 付瑞,李剑宇,王笳辉,等.面向领域知识图谱的实体关系联合抽取 [J].华东师范大学学报 (自然科学
版),2021(5):24 - 36.
[31] WANGP,YANGA,MEN R,etal.Ofa:Unifyingarchitectures,tasks,andmodalitiesthroughasimplese
quence - to - sequencelearningframework[C]??InternationalConferenceonMachineLearning.PMLR,2022.
[32] 李可悦,陈轶,牛少彰.基于 BERT的社交电商文本分类算法[J].计算机科学,2021,48(2):87 - 92.
[33] LANDISJR,KOCHGG.Anapplicationofhierarchicalkappa - typestatisticsintheassessmentofmajorityagree
mentamongmultipleobservers [J].Biometrics,1977:363 - 374.
[34] JIY,SUNA,ZHANGJ,etal.A criticalstudyondataleakageinrecommendersystem offlineevaluation[J].
ACM TransactionsonInformationSystems ,2023,41(3):1 - 27.
[35] JIANGK,PRADEEPR,LINJ.ExploringlistwiseevidencereasoningwithT5forfactverification[C]??Proceed
ingsofthe59thAnnualMeetingoftheAssociationforComputationalLinguisticsandthe11thInternationalJoint
ConferenceonNaturalLanguageProcessing (Volume2:ShortPapers).2021.
Intelligentanalysisandjointextractionofrescueentitiesand
relationshipsinwaterprojecttexts
1,2
1
3
1
1
YANGYangrui,ZHUYaping,LIUXuemei ,CHENSisi,LIHuimin
(1.SchoolofInformationEngineering,NorthChinaUniversityofWaterResourcesandElectricPower,Zhengzhou 450000,China;
2.CollaborativeInnovationCenterforEfficientUtilizationofWaterResources,Zhengzhou 450000,China;
3.SchoolofWaterConservancy,NorthChinaUniversityofWaterResourcesandElectricPower,Zhengzhou 450000,China)
Abstract:Waterprojectrescuemeasuresareanimportantpartoffloodpreventionemergencyplan.Thisarticle
aimstouseinformationextractiontechnologytoextractwaterprojectrescueknowledgefrom variousunstructured
textsources ,andtransform itintoatriplestructureof〈entity,relationship,entity〉,andprovidestructured
knowledgesupportforintelligentgenerationofemergencyplans.Theheterogeneouswaterprojectrescueentityex
tractionandrelationshipextractiontasksareconsideredassequence - to - sequencegenerationtasks ,andwaterpro
jectrescueentitiesandrelationshipsjointextraction (WRERJE)frameworkbasedonlargelanguagemodelsispro
posed.WRERJEisamultitaskingframeworkforbothentityextractionandrelationshipextraction ,whichusesdy
namicpromptstoguideT5forjointextractionofentitiesandrelationships.Thetextdataaugmentationmethodspe
cifictothefieldofwaterprojectrescueisstudied ,andonthebasisofthepreliminaryfine - tuningofWRERJEby
usingasmallnumberoflabeledsamples ,WRERJEisfurtherfine - tunedbyusingmorevaguelydescribedbutcor
rectlylabeleddataareobtainedbydataaugmentationmethod,improvingitsperformanceforextractingwaterproject
rescueentitiesandrelationships.TheperformanceofWRERJEisevaluatedexperimentally ,andtheresultsshow
thatWRERJEshowshighextractionperformanceinthetaskofwaterprojectrescueentityextractionandrelationship
extraction(F1valuesofentityandrelationshipreach78.42% and78.22%,respectively),whichverifiestheef
fectivenessofdynamicpromptandjointextractionmethods.
Keywords:waterprojectrescue;emergencyplan;informationextraction; dynamicprompt; jointextraction;
textdataaugmentation
(责任编辑:于福亮)
— 8 2 —
8