Upgrade to Pro — share decks privately, control downloads, hide ads and more …

생성형AI를 활용한 개체명인식(NER)

kakao
November 01, 2024

생성형AI를 활용한 개체명인식(NER)

#NER #LLM #NLP

비정형데이터에서 중요한 정보를 추출하기 위해서는 패턴 뿐만아니라 전문적인 지식과 의미 또한 고려해야 합니다. 개체명인식(NER)은 정보를 추출하는 가장 첫 과정이면서 동시에 가장 중요한 과정입니다.
병원 내 비정형데이터(영상검사기록지, 병리검사기록지 등)에 생성형AI를 활용해 NER하는 과정을 통해 잠들어있던 중요한 의료 정보를 가치있게 만들어가는 방법을 공유합니다.

발표자 : brown.sung
카카오헬스케어에서 AI 프로젝트를 총괄하고 있습니다.

kakao

November 01, 2024
Tweet

More Decks by kakao

Other Decks in Programming

Transcript

  1. ࠺੿ഋؘ੉ఠ <'JOEJOH> ೨ઙ''%( Ѩࢎߑߨ''%(ܳ*7JOKFDUJPOറदр૩1&5$5੹न৔࢚ਸ୫৔ೞ৓ਵݴ ೙ਃীٮۄ୶о୫৔ਸ पदೣ ੐࢚ࣗѼ-UBOLMFQBJO PDDBTJPOBMTXFMMJOH GV #POFBOEKPJOU

    -FGUBOLMFKPJOUীࢲGPDBMVQUBLF੉ҙ଴غݴ ੉੹ࠁ׮ࣗѼ੉ঈചؽ নஏTBDSPJMJBDKPJOUীࢲب࠺؀ட੸ੋGPDBMVQUBLF੉੓ਵݴ -ଃ੉3ଃࠁ׮؊௾VQUBLFਸࠁ੐ -3  3JHIUIJQKPJOUীࢲNJMEVQUBLFૐоо੓ਵݴ MFGUTIPVMEFSKPJOUীࢲبIPUVQUBLF੉ҙ଴ؽনଃ LOFFীࢲبNVMUJQMFJODSFBTFEVQUBLF੉ࠁ੉ݴ ౠ൤MFGUTJEFীࢲ؊QSPNJOFOUೣ -FGUXSJTU৬IBOEKPJOU੄VQUBLFبૐоغয੓ਵݴ ੉חঌ۰૓PTUFPBSUISJUJTী੄ೠ߸ച۽ࠁ੐ •OPTJHOJGJDBOUJOUFSWBMDIBOHF 5WFSUFCSBMCPEZীࢲJODSFBTFEVQUBLF੉੓ਵݴ .3*ীࢲഛੋػEJTLEFHFOFSBUJPOী੄ೠ ࣗѼੌоמࢿ੉֫਺ •TMJHIUMZJODSFBTFEVQUBLFDPNQBSFEUPQSFWJPVTTUVEZ ୶о۽ $TQJOF 5 -ࠗਤীࢲب׮ࣻ੄JODSFBTFEVQUBLF੉ഛੋؽ ࠄ ੗ܐীחࢤࢿഋ"*Ѿҗޛ੉ੌࠗನೣغয੓णפ׮
  2. 3FBM8PSME&WJEFODF पࢎਊૐѢ 3FBM8PSME%BUB 38% 1IBSNBEBUB 3$5 PCTFSWBUJPOBM &MFDUSPOJDNFEJDBM SFDPSET %FWJDF

    EBUB .PSUBMJUZ PUIFS SFHJTUSJFT )PTQJUBMWJTJUT TFSWJDFEFUBJMT -BCUFTUT QBUIPMPHZSFTVMUT $MBJNT EBUB 4PDJBM NFEJB $POTVNFS EBUB .FBOJOHGVM RVFTUJPOT 'JUGPS QVSQPTFEBUB BOBMZUJDT 3FBM8PSME &WJEFODF 3FBM8PSME&WJEFODF 38& 3FBM8PSME&WJEFODFJTUIFDMJOJDBMFWJEFODF SFHBSEJOHUIFVTBHFBOEQPUFOUJBMCFOFGJUTPSSJTLT PGBNFEJDBMQSPEVDUEFSJWFEGSPNBOBMZTJTPG38%
  3. /&3 /BNFE&OUJUZ3FDPHOJUJPO ੉ۆ ddx change over time anatomical location anatomical

    location name name size size "CPVUNN SFTJEVBMSFDVSSFOUBOFVSZTNBU UIFDPJMFNCPMJ[FESJHIU*$"CJGVSDBUJPO /PDIBOHF PGB NN PVUQPVDIJOHMFTJPOBU UIFMFGUQBSBDMJOPJE *$"  30BOFVSZTNBMDIBOHF
  4. /&3 /BNFE&OUJUZ3FDPHOJUJPO ੉ۆ /P &OUJUZ 3FTVMU  /BNF SFTJEVBMSFDVSSFOBU BOFVSZTN

    "OBUPNJDBM MPDBUJPO UIFDPJMFNCPMJ[FESJHIU*$" CJGVSDBUJPO 4J[F "CPVUNN  /BNF PVUQPVDIJOHMFTJPO "OBUPNJDBM MPDBUJPO UIFMFGUQBSBDMJOPJE *$" 4J[F NN $IBOHFPWFS UJNF /PDIBOHF EEY 30BOFVSZTNBMDIBOHF j j j ੗োয୊ܻ /-1 ܳਤೠо੢ӝୡ੸੉Ҋ೙ࣻ੸ੋ ੘স  ੿ഛೠ/&3୊ܻחੋр੄঱যܳӝ҅о ੋधೞӝਤೠױ҅۽ ױযೞաೞա੄੄޷ܳ ੋधೞѱೞחҗ੿  ѐ୓ী؀ೠ੄޷৬Ӓ੄޷ী೧׼ೞחױযܳ ష௾ਵ۽ٜ݅য୊ܻ
  5. --.ӝ߈/&3 ҳ࠙ IVNBOMBCPSߑध --.ӝ߈ /&3 ࠺Ү ୊ܻ ࣘب ୡѤ Ѥੌ

    ୊ܻ ӝળ _ୡѤ ড _ߓ )(16ѐࣻ _੢ ӝળ ࣗਃदр ݅Ѥ ӝળ ੌ ੌ ড ੉ೞ .BYୡѤ ӝળ ࠺ਊ ݅Ѥ ӝળ ড ୌߔ݅ਗ ড ߔ݅ਗ ড ੉ೞ ੋѤ࠺ ਘ ݅ਗݺ ӝળ ࢿמ ಣӐ  IVNBOFSSPSӝળ ੉࢚ 'TDPSFӝળ
  6. --.ӝ߈/&3GSBNFXPSL 'JMFTFSWFS 8PSLTUBUJPO --.TFSWFS 4BNQMFS -BCFM 4UVEJP 1SPNQU .BOBHFS 1FSGPSNB

    ODF FWBMVBUJPO EBUBTFU%# USBJOJOHEBUB WBMJEBUJPOEBUB 'PVOEBUJPO .PEFM 'JOF5VOJOH 'JOFUVOFE .PEFM *OGFSFODF SBXEBUB SFTVMU EBUB BOOPUBUJPO ؘ੉ఠ ഐ୹ ੽Ӕాઁ JOGFSFODF ؘ੉ఠ ഐ୹ JOGFSFODF Ѿҗ ੸੤
  7. --.ӝ߈/&3GSBNFXPSL ӝ۾૑ ࢶ੿ా೤ ؘ੉ఠ ੿੄ࢲ ೐܁೐౟ ےؒ ࢠ೒݂ ؘ੉ఠ ੹୊ܻ

    BOOPUBUJPO • GFXTIPU߂ పझ౟ ਊ ѐ • 5SBJOJOHਊ ѐ • 5FTUਊ ѐ ؘ੉ఠ ੿੄ࢲ ѐࢶ ೐܁೐౟ ѐࢶ 'JOFUVOJOH ୶ۿ ࢿמ Ҋبച प೷ ߈ࠂ ૓೯ ળ࠺ ױ҅ য֢ప੉࣌ ױ҅ ࢿמҊبച ױ҅ ୶ۿ ױ҅
  8. &YQFSJNFOU प೷੉ۆ ా҅੸੽Ӕ оࢸਸࢸ੿ೞҊ ೧׼оࢸ੉ଵੋ૑Ѣ૙ੋ૑ా҅੸ਬ੄ࢿਵ۽ഛੋೞח੺ର ѐߊҙ੼ "#UFTU --.਷৵प೷੉೙ਃೠо à ୭੸੄ઑ೤ਸ଺חѪ੉୭Ҋ੄ࢿמ

    ࢤࢿഋ"*ীࣻ೯ೞחসޖীٮۄઑ੿੉೙ਃೠ߸ࣻ߂Ӓઑ೤੉ޖೠীоө਑ ӝ۾૑߹۽৘࢚ೞ૑ޅೠ੉गٜ੉૑ࣘ੸ਵ۽ߊࢤೡࣻ੓਺à ౠ൤ ജп IBMMVDJOBUJPO 'JOFUVOJOHਸೞݶUSBEFPGG੉गоߊࢤm ਗېੜೞ؍୶ۿמ۱੉ծই૑חޙઁ 모델 선정, 모델 파라미터 조정, few-shot 개수, few-shot 선정, fine-tuning 데이터 건수, fine-tuning 데이터 내용, fine-tuning 시 few-shot/prompt 반영 여부 등등
  9. &YQFSJNFOU അ੤ө૑ߋഃմࢎप Ÿ 'JOFUVOOJOHXJUIOPQSPNQU  QSPNQUহ੉GJOFUVOOJOHೞחѪ੉ࡅܰҊ ࢿמҊبച੄оמࢿ੉֫ই૓׮  ౠ੿੘স૑ஜਸGJOFUVOJOHীನೣदఃחѪী؀ೠ׮নೠ੄Ѽઓ੤ Ÿ

    *OGFSFODFXJUI'FXTIPUCBTFE'JOFUVOFE.PEFM  GJOFUVOOJOHػݽ؛ীGFXTIPU݅֍যبળࣻೠࢿמ਷ࠁ੢ػ׮ Ÿ 2VBOUJUZY 2VBMJUZ  GJOFUVOOJOHীࢎਊೞחؘ੉ఠѐࣻ৬࠺۹ೞৈࢿמೱ࢚੉׸ࠁغ૑חঋח׮  ಁఢ੄׮নࢿ ղਊ੄ࠂ੟ࢿীٮۄ୭੸੄GJOFUVOJOHؘ੉ఠѐࣻо੓חо
  10. ݃ޖܻ ೾झாয࠙ঠחؘ੉ఠ ഝਊ 38& ੉оࣘ઺ ӝ۾૑חݒ਋઺ਃೠ ജ੗੿ࠁ੄ࠁҊ ઺ਃೞҊ೙ࣻ੸ੋ੿ࠁܳ ੜ୶୹ೞחѪ੉/&3੄೨ब /&3ীࢲ--.਷USZ

    FSSPS੄߈ࠂ ੗زചೞח җ੿੉઺ਃ ѐߊ੗ بݫੋ੹ޙо  ੄ࢎѾ੿੗р੄ࣗా઺ਃ খਵ۽೧ঠೡ݆਷Ѫٜj ࢿמ ߧਊࢿ ١١
  11. 2"