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

論文紹介: Corpora Generation for Grammatical Error Correction

論文紹介: Corpora Generation for Grammatical Error Correction

- 首都大 小町研 春の最先端論文紹介 2019
- Corpora Generation for Grammatical Error Correction. Lichtarge et al., NAACL2019 を紹介
- paper URL: https://www.aclweb.org/anthology/N19-1333.pdf

Satoru Katsumata

December 10, 2023
Tweet

More Decks by Satoru Katsumata

Other Decks in Research

Transcript

  1. ֓ཁ w จ๏ޡΓగਖ਼͸ޡΓจˠగਖ਼จͷ຋༁λεΫͱͯ͠
 ѻΘΕΔ͜ͱ͕ଟ͍
 ˠ࢖༻Ͱ͖Δֶशऀؔ࿈ͷର༁ίʔύεʹ͸ݶΓ͕͋Δ w 8JLJQFEJBσʔλΛ΋ͱʹֶशσʔλΛ૿΍͢
 8JLJQFEJBฤूཤྺΛ࢖༻͢Δ
 ಛʹɺ࠷௿ݶͷϑΟϧλϦϯάΛࢪͯ͠
 େྔͷɺ࣮ࡍͷޡΓͰߏ੒͞Ε͍ͯΔ͕Ԛ͍σʔλΛ֫ಘ͢Δ


    3PVOEUSJQ5SBOTMBUJPOʢ355࠶຋༁ʣΛ࢖༻͢Δ
 UHU8JLJQFEJBநग़จTSDUHUΛ࠶຋༁ͨ͠΋ͷ
 ຋༁༝དྷͰ͸͋Δ͕ɺ8JLJQFEJB΄ͲԚ͘ͳ͍σʔλΛ֫ಘ͢Δ w QSFUSBJO࡞੒ͨ͠σʔλpOFUVOF-BOH
 ˠ$P/--ͱ+'-&(ͰධՁɺ1#4.5ͱ/.5ͷϋΠϒϦουϞσϧͱ
 Ͳ͍ͬ͜ͷੑೳ !2
  2. 8JLJQFEJBฤूཤྺ༝དྷͷ
 σʔλ࡞੒ w ൈ͖ग़͢લͷॲཧ
 .CҎ্ͷϖʔδ͸ল͘
 ϖʔδશͯͷฤूΛ࢖͏ͷͰ͸ͳ͘ɺର਺εέʔϧͰநग़͢Δ
 O͸ϖʔδͷUPUBMSFWJTJPO਺ w ൈ͖ग़ͨ͠ޙͷॲཧˠDPSSVQUJPOBOEpMUSBUJPO
 DPSSVQUJPO


    TSDଆʹDIBSͷ֬཰ͰTQFMMJOHFSSPSΛೖΕΔ
 ʢ࡟আɺૠೖɺஔ׵ɺྡ઀ೖΕସ͑ʣ
 pMUSBUJPO
 ࠷େจ௕XPSEQJFDF
 TSDUHUͷ৔߹΄ͱΜͲʢʣࣺͯΔ
 ˠ࠷ऴతʹSFWJTJPOEBUBTFUͷ͕ಉҰจର w ͜͜ͷ࣮૷͸UFOTPSUFOTPSͷEBUB@HFOFSBUPSTʹೖͬͯΔ log1.5 (n) !3
  3. 8JLJQFEJBฤूཤྺ༝དྷͷ
 σʔλ࡞੒ w ൈ͖ग़͢લͷॲཧ
 .CҎ্ͷϖʔδ͸ল͘
 ϖʔδશͯͷฤूΛ࢖͏ͷͰ͸ͳ͘ɺର਺εέʔϧͰநग़͢Δ
 O͸ϖʔδͷUPUBMSFWJTJPO਺ w ൈ͖ग़ͨ͠ޙͷॲཧˠDPSSVQUJPOBOEpMUSBUJPO
 DPSSVQUJPO


    TSDଆʹDIBSͷ֬཰ͰTQFMMJOHFSSPSΛೖΕΔ
 ʢ࡟আɺૠೖɺஔ׵ɺྡ઀ೖΕସ͑ʣ
 pMUSBUJPO
 ࠷େจ௕XPSEQJFDF
 TSDUHUͷ৔߹΄ͱΜͲʢʣࣺͯΔ
 ˠ࠷ऴతʹSFWJTJPOEBUBTFUͷ͕ಉҰจର log1.5 (n) !4
  4. ࠶຋༁༝དྷͷσʔλ࡞੒ w 8JLJQFEJBSFWJTJPOEBUBͰࣺͯͨTSDUHUͷσʔλʹରͯ͠ɺ
 ҎԼͷΑ͏ʹ3PVOEUSJQUSBOTMBUJPOΛߦ͍ٖࣅTSDΛ࡞੒͢Δ
 USBOTMBUJPONPEFM(/.5 8VFUBM   w ੜ੒ͨ͠ޙͷॲཧˠDPSSVQUJPO


    TSDUHUͳΔࣄྫ͕΄΅ͳ͍ˠશମͷʹͳΔΑ͏ʹ૿΍͢
 ͷ֬཰ͰTQFMMJOHFSSPSʢ࡟আɺૠೖɺྡ઀సஔʣΛಋೖ
 SFWJTJPOEBUBʹଘࡏ͢Δగਖ਼લޙͷ૊Έ߹ΘͤΛ൓өͤ͞Δ
 ˠSFWJTJPOEBUBͱಉ͡αΠζʹͳΔΑ͏ʹͨ͠ lang ∈ {French, German, Japanese, Russian} UHUTFOU
 FOXJLJ USBOTMBUFETFOU
 MBOHXJLJ TSDTFOU
 FOXJLJ USBOTMBUF
 FOMBOH USBOTMBUF
 MBOHFO !5
  5. ࠶຋༁༝དྷͷσʔλ࡞੒ w 8JLJQFEJBSFWJTJPOEBUBͰࣺͯͨTSDUHUͷσʔλʹରͯ͠ɺ
 ҎԼͷΑ͏ʹ3PVOEUSJQUSBOTMBUJPOΛߦ͍ٖࣅTSDΛ࡞੒͢Δ
 USBOTMBUJPONPEFM(/.5 8VFUBM   w ੜ੒ͨ͠ޙͷॲཧˠDPSSVQUJPO


    TSDUHUͳΔࣄྫ͕΄΅ͳ͍ˠશମͷʹͳΔΑ͏ʹ૿΍͢
 ͷ֬཰ͰTQFMMJOHFSSPSʢ࡟আɺૠೖɺྡ઀సஔʣΛಋೖ
 SFWJTJPOEBUBʹଘࡏ͢Δగਖ਼લޙͷ૊Έ߹ΘͤΛ൓өͤ͞Δ
 ˠSFWJTJPOEBUBͱಉ͡αΠζʹͳΔΑ͏ʹͨ͠ lang ∈ {French, German, Japanese, Russian} UHUTFOU
 FOXJLJ USBOTMBUFETFOU
 MBOHXJLJ TSDTFOU
 FOXJLJ USBOTMBUF
 FOMBOH USBOTMBUF
 MBOHFO  8JLJSFWJTJPO͔Β୯ޠҎ಺ͷFEJUΛݟ͚ͭΔ  ରԠ͢ΔPSJHJOBMͱSFWJTFͷ૊Έ߹ΘͤΛΧ΢ϯτɺ
 ҎԼͷΑ͏ʹPSJHJOBMQISBTFͷ֬཰ΛٻΊΔ  ͜ͷ֬཰Λ༻͍ͯTSDଆʹޡΓʢPSJHJOBMQISBTFʣΛೖΕ͍ͯ͘ P(original|revised) = C(original, revised) C(revised) !6
  6. ࣮ݧઃఆ w Ϟσϧ5SBOTGPSNFS 5FOTPS5FOTPSͷUSBOTGPSNFS@DMFBO@CJH@UQV
 PQUJNJ[FS"EBGBDUPSMBZFS IFBE FNCFEEJN GFFEGPSXBSE 
 CFBN

    w XPSEQJFDF࢖༻ w ֶश8JLJQFEJBSFWJTJPOPS355ˠ-BOHEBUB
 QSFUSBJOFQPDI CBUDITJ[F XPSEQJFDF 
 MFBSOJOHSBUFpSTU TUFQˠͦͷ͋ͱ͸ˠ
 pOFUVOFFQPDIMFBSOJOHSBUFº?
 ʢͨͩ͠ผͰͱ͓͍ͬͯͨ-BOHʁΛEFWͱͯ͠ऩଋ൑ఆʣ w ධՁσʔλ$P/--BOE+'-&(ʢ+'-&(ͷΈTQFMMDIFDL༗ʣ w JUFSBUJWFEFDPEJOH
 UISFTIPME+'-&(͸EFWͰܾఆ͢Δ
 $P/--͸$P/--USBJOJOHTFU /6$-& ͔Β$P/--EFWTFUͱಉ͡గਖ਼จɺ ඇగਖ਼จͷׂ߹ʹͳΔΑ͏ʹTBNQMFͨ͠σʔλͰܾΊΔʢ$P/--USBJOJOHTVCTFUʣ
 JUFSBUJPO$P/--USBJOJOHTVCTFUͰܾఆɺ+'-&(΋ಉ͡஋Λ࢖͍·Θͯ͠Δʁ w ͦͷ΄͔
 TQFMMDIFDLFS+'-&(ͷΈ lrstep = lrstep−1 step !10
  7. ଞͷγεςϜͱͷൺֱʙ #FTUTJOHMFNPEFM8JLJQFEJBSFWJTJPO.BYFEJU
 #FTU&OTFNCMFSFWJTJPO 355NPEFM
 QVCMJDEBUBͷΈ࢖༻͍ͯ͠ΔϞσϧؒͰͷൺֱ  $IPMMBNQBUUBOE/H  $//FODEFD 

    +VOD[ZT%PXNVOUFUBM  5SBOTGPSNFSXJUIMPXSFTPVSDFUJQT  (SVOELJFXJD[BOE+VOD[ZT%PXNVOU  CJ(36FODEFDXJUIBUUFOUJPOBOE1#4.5 !14
  8. উखʹ
 ଞͷQSFUSBJOϞσϧͱൺֱʙ w /""$-࠾୒࿦จ*NQSPWJOH(SBNNBUJDBM&SSPS$PSSFDUJPOWJB1SF5SBJOJOHB $PQZ"VHNFOUFE"SDIJUFDUVSFXJUI6OMBCFMFE%BUB;IBPFUBM w ͜Ε͸ࣗಈཁ໿ͰྲྀߦͬͯΔ1PJOUFS(FOFSBUPSΛ(&$ʹ΍ͬͨ࿩ͰɺNPEFMͷQSFUSBJO ʹ0OF#JMMJPO8PSE#FODINBSLʢ.TFOUTʣΛجʹٖͨ͠ࣅσʔλΛ࢖༻͍ͯ͠Δ w ۩ମతʹ͸0OF#JMMJPOͷPSJHJOBMΛUHUʹɺ


    PSJHJOBMʹରͯ͠࡟আɺૠೖɺஔ׵ɺೖΕସ͑ૢ࡞Λద౰ʹߦͬͨ΋ͷΛTSDͱͯ͠ EFOPJTJOHBVUPFODPEFSͷؾ࣋ͪͰQSFUSBJO͍ͯ͠Δ w ঺հ͍ͯ͠Δ࿦จͱͷҧ͍NPEFMΛ޻෉͠ɺֶशσʔλ͕গྔʹͳͬͨ !15 ;IBPFUBMͷQSFUSBJOSFTVMU 3FWJTJPOQSFUSBJOͷ݁Ռ
  9. ࢀߟจݙ w *NQSPWJOH(SBNNBUJDBM&SSPS$PSSFDUJPOWJB1SF5SBJOJOHB$PQZ"VHNFOUFE"SDIJUFDUVSF XJUI6OMBCFMFE%BUB;IBPFUBM /""$- w (PPHMFT/FVSBM.BDIJOF5SBOTMBUJPO4ZTUFN#SJEHJOHUIF(BQCFUXFFO)VNBOBOE .BDIJOF5SBOTMBUJPO8VFUBM BS9JW w

    5IF8JL&E&SSPS$PSQVT"$PSQVTPG$PSSFDUJWF8JLJQFEJB&EJUTBOE*UT"QQMJDBUJPOUP (SBNNBUJDBM&SSPS$PSSFDUJPO(SVOELJFXJD[BOE+VOD[ZT%PXNVOU *OUFSOBUJPOBM $POGFSFODFPO/BUVSBM-BOHVBHF1SPDFTTJOH w ".VMUJMBZFS$POWPMVUJPOBM&ODPEFS%FDPEFS/FVSBM/FUXPSLGPS(SBNNBUJDBM&SSPS $PSSFDUJPO$IPMMBNQBUUBOE/H """* w "QQSPBDIJOH/FVSBM(SBNNBUJDBM&SSPS$PSSFDUJPOBTB-PX3FTPVSDF.BDIJOF 5SBOTMBUJPO5BTL+VOD[ZT%PXNVOUFUBM /""$- w /FBS)VNBO-FWFM1FSGPSNBODFJO(SBNNBUJDBM&SSPS$PSSFDUJPOXJUI)ZCSJE.BDIJOF 5SBOTMBUJPO(SVOELJFXJD[BOE+VOD[ZT%PXNVOU /""$- !18