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

Research Paper Introduction in IR Reading 2020 Fall

Takuya Asano
October 31, 2020

Research Paper Introduction in IR Reading 2020 Fall

1. Finding the Best of Both Worlds:
Faster and More Robust Top-k Document Retrieval
(SIGIR 2020)
 https://dl.acm.org/doi/abs/10.1145/3397271.3401076
2. Efficient and Effective Query Auto-Completion
(SIGIR 2020) https://dl.acm.org/doi/abs/10.1145/3397271.3401432

IR Reading 2020 Fall https://sigir.jp/post/2020-10-31-irreading_2020fall/

Takuya Asano

October 31, 2020
Tweet

More Decks by Takuya Asano

Other Decks in Research

Transcript

  1. 5BLVZB"TBOP ઙ໺୎໵

    .FSDBSJ *OD
    'JOEJOHUIF#FTUPG#PUI8PSMET
    'BTUFSBOE.PSF3PCVTU5PQL%PDVNFOU3FUSJFWBM
    4*(*3

    &GGJDJFOUBOE&GGFDUJWF2VFSZ"VUP$PNQMFUJPO
    4*(*3


    *33FBEJOH'BMM
    0DUPCFS

    View Slide

  2. 'JOEJOHUIF#FTUPG#PUI8PSMET
    'BTUFSBOE.PSF3PCVTU5PQL%PDVNFOU3FUSJFWBM 4*(*3

    0NBS,IBUUBC 4UBOGPSE6OJWFSTJUZ

    .PIBNNBE)BNNPVE $BSOFHJF.FMMPO6OJWFSTJUZJO2BUBS

    5BNFS&MTBZFE 2BUBS6OJWFSTJUZ

    w ୅දతͳUPQLSFUSJFWBMख๏ͷ଎౓Λɺ༷ʑͳϥϯΩϯάϞσϧͰแׅతʹධՁ
    w 8"/%ͱ.BY4DPSFͷ͍͍ͱ͜औΓΛͨ͠৽ͨͳख๏ɺ-B[Z#.ΛఏҊ
    w ΄ͱΜͲͷ࣮ݧઃఆͰɺఏҊख๏-B[Z#.͕࠷଎Λୡ੒

    View Slide

  3. 03ΫΤϦͷ5PQL3FUSJFWBM
    &BSMZ5FSNJOBUJPO
    w "/%ΫΤϦ͸సஔΠϯσοΫεΛ࢖ͬͯߴ଎ʹॲཧͰ͖Δ
    w 03ΫΤϦ͕໰୊
    w ߴ଎Խ͢Δํ๏͸͍Ζ͍Ζ͋Δ͕ɺ&BSMZ5FSNJOBUJPOʹ஫໨
    w ਖ਼֬ͳUPQLΛಘΒΕΔ
    w 5PQLʹೖΒͳ͍͜ͱ͕Θ͔ͬͨจॻ͸είΞܭࢉ͠ͳ͍
    w ୅දతͳͭͷઓུʹ8"/%ͱ.BY4DPSF͕͋Δ
    w λʔϜ͝ͱͷείΞͷ࠷େ஋ NBYTDPSF
    Λϝλσʔλͱͯ͠ར༻

    View Slide

  4. 8"/%ΞϧΰϦζϜ
    03ΫΤϦͷ5PQL3FUSJFWBM
    $POTUBOUJOPT%JNPQPVMPT 4FSHFZ/FQPNOZBDIJZ BOE5PSTUFO4VFM0QUJNJ[JOHUPQLEPDVNFOU
    SFUSJFWBMTUSBUFHJFTGPSCMPDLNBYJOEFYFT*O84%.
    ݱࡏͷUPQLͷείΞ
    ϐϘοτλʔϜͷબ୒
    w λʔϜΛΧʔιϧҐஔͷঢॱͰιʔτ
    w NBYTDPSFΛྦྷੵ͍͖ͯ͠ɺВΛ௒͑ͨλʔϜΛϐϘοτͱͯ͠
    બ୒͢Δ
    ΧʔιϧҐஔ߹Θͤ
    w ϐϘοτΑΓ্ʹ͋ΔλʔϜͷΧʔιϧҐஔΛɺϐϘοτͷΧʔ
    ιϧҐஔ·ͰਐΊΔ
    w ϐϘοτͷΧʔιϧҐஔΛ௨Γա͗ͨΒɺͦͷจॻ͸εΩοϓ
    ʢʹ໭ͬͯϐϘοτΛ࠶બ୒ʣ
    จॻͷείΞܭࢉ
    w จॻείΞΛ׬શʹܭࢉ
    w ВΛӽ͑ͨ৔߹͸UPQLʹೖΕɺВͷ஋Λߋ৽
    λʔϜ͝ͱͷείΞͷ࠷େ஋
    ʢϝλσʔλʣ
    ϐϘοτͷΧʔιϧҐஔ

    View Slide

  5. .BY4DPSFΞϧΰϦζϜ
    03ΫΤϦͷ5PQL3FUSJFWBM
    $POTUBOUJOPT%JNPQPVMPT 4FSHFZ/FQPNOZBDIJZ BOE5PSTUFO4VFM0QUJNJ[JOHUPQLEPDVNFOU
    SFUSJFWBMTUSBUFHJFTGPSCMPDLNBYJOEFYFT*O84%.
    ݱࡏͷUPQLͷείΞ
    λʔϜ͝ͱͷείΞͷ࠷େ஋
    ʢϝλσʔλʣ
    FTTFOUJBMλʔϜͱOPOFTTFOUJBMλʔϜʹ෼͚Δ
    w ্͔ΒNBYTDPSFΛྦྷੵ͍͖ͯ͠ɺВΛ௒͑ͳ͍λʔϜू߹ΛOPO
    FTTFOUJBMλʔϜɺͦΕҎ֎ΛFTTFOUJBMλʔϜͱ͢Δ
    w FTTFOUJBMλʔϜͷ͏ͪɺগͳ͘ͱ΋ͭ͸ؚΜͰ͍ͳ͍ͱUPQLʹೖΕͳ͍
    FTTFOUJBMλʔϜʹ͍ͭͯΫΤϦॲཧ
    w FTTFOUJBMλʔϜͷதͰείΞͷ߹ܭ஋Λܭࢉ
    w OPOFTTFOUJBMλʔϜͷNBYTDPSFͷ߹ܭ QSFpYTVN
    ͱ଍ͯ͠Вʹಧ͔
    ͳ͚Ε͹ɺ͜͜Ͱଧͪ੾ͬͯΧʔιϧΛਐΊΔ
    w OPOFTTFUOBJMλʔϜͷϙεςΟϯάϦετ΋ݟͯείΞܭࢉ͠ɺ෦෼είΞ
    ͕ВΛԼճͬͨΒଧͪ੾ͬͯΧʔιϧΛਐΊΔ
    จॻͷείΞܭࢉ
    w จॻείΞΛ׬શʹܭࢉ͠ɺВΛӽ͑ͨ৔߹͸UPQLʹೖΕɺВͷ஋Λߋ৽
    w ඞཁ͕͋Ε͹FTTFOUJBMλʔϜΛOPOFTTFOUJBMʹམͱ͢
    λʔϜ͸
    NBYTDPSFͰ
    ιʔτࡁΈ

    View Slide

  6. 8"/%ϕʔεͷΞϧΰϦζϜ
    ֓ཁ
    w 8"/% 8FBL"/%

    w ϐϘοτબ୒ʹΑΔ&BSMZ5FSNJOBUJPO
    w #.8 #MPDL.BY8"/%

    w ϙεςΟϯάϦετΛϒϩοΫԽ
    w -#.8 -PDBM#.8

    w ϐϘοτબ୒ʹϩʔΧϧͳʢϒϩοΫ͝ͱͷʣNBYTDPSFΛར༻
    w %#.8 %PD*%PSJFOUFE#.8

    w ϒϩοΫΛݻఆ௕Ͱ͸ͳ͘ɺจॻ*%ͷݻఆ௕ͷΠϯλʔόϧͰ෼͚Δ
    w 7#.8 7BSJBCMF#.8

    w ϒϩοΫΛՄมαΠζʹɻ֤ϒϩοΫαΠζ͸ಈతܭը๏Ͱ࠷దԽ

    View Slide

  7. ධՁ
    w ΄ͱΜͲશͯͷઃఆͰ%#.8
    %PD*%PSJFOUFE#MPDL.BY
    8"/%
    ͕࠷଎
    w %PD*%PSJFOUFE#MPDLΛ࢖͏
    ΞϧΰϦζϜ
    w -.%JSϥϯΩϯάϞσϧͰ
    -#.8 7#.8ʹΘ͔ͣʹෛ
    ͚͍ͯΔ͚ͩ
    8"/%ϕʔεͷΞϧΰϦζϜ

    View Slide

  8. .BY4DPSFϕʔεͷΞϧΰϦζϜ
    ֓ཁ
    w .BY4DPSF
    w λʔϜΛFTTFOUJBMOPOFTTFOUJBMʹ෼͚ͯ&BSMZ5FSNJOBUJPO
    w #.. #MPDL.BY.BY4DPSF

    w .BY4DPSFΛϒϩοΫԽ
    w -#.. -PDBM#..

    w ϩʔΧϧͷʢϒϩοΫ͝ͱͷʣNBYTDPSFΛར༻
    w *#.. *OUFSWBMCBTFE#..

    w จॻ*%ͷൣғΛϒϩοΫ͝ͱͷNBYTDPSFΛ࢖ͬͯ෼ׂ
    w %#.. %PD*%CBTFE#..

    w %#.8ͱಉ༷ɺ%PD*%CBTFE#MPDLΛ࢖͏

    View Slide

  9. ධՁ
    w .BY4DPSFϕʔεͷख๏ͱൺֱ
    ͯ͠΋ɺ΄ͱΜͲશͯͷઃఆͰ
    %#.8͕࠷଎
    w -.%JSϥϯΩϯάϞσϧͰΘ
    ͔ͣʹ%#.. *#..ʹෛ͚
    ͍ͯΔ͚ͩ
    .BY4DPSFϕʔεͷΞϧΰϦζϜ

    View Slide

  10. ֓ཁ
    w "NPSUJ[FE1JWPU'JMUFSJOH
    w ϒϩοΫߏ଄ͱͯ͠%PD*%PSJFOUFE
    #MPDL.BYΛ࠾༻
    w #BMBODFE1SVOJOH
    w 8"/%ϕʔεͷΫΤϦॲཧ
    w -PDBM5FSN1BSUJUJPOJOH
    w .BY4DPSFͱಉ༷ɺλʔϜΛ
    FTTFOUJBMͱOPOFTTFOUJBMʹ෼͚Δ
    ఏҊख๏-B[Z#.
    8"/%ϕʔε
    ͷϑΟϧλ
    .BY4DPSFϕʔε
    ͷϑΟϧλ

    View Slide

  11. 405"Ͱ͋Δ%#.8ͱൺֱͯ͠ɺ-B[Z#.͸Ͳ͏͔ʁ
    ΋͠ੑೳ্͕ճ͍ͬͯΔͳΒɺͦΕ͸Ͳ͔͜Βདྷ͍ͯΔͷ͔ʁ
    32

    View Slide

  12. 405"WT-B[Z#.
    w ΫΤϦͷϨΠςϯγΛൺֱ
    w LL
    w .FBO5BJM ύʔηϯλΠϧ఺

    w ͢΂ͯͷύλʔϯͰ-B[Z#.͕༏Ґ
    w ϥϯΩϯάϞσϧฏۉʢ%#.8WT-B[Z#.ʣ
    w .FBOഒߴ଎
    w 5BJMഒߴ଎
    ϝΠϯϦβϧτ

    View Slide

  13. -B[Z#.ͷߴ଎Խ͸ͲͷΑ͏ʹಘΒΕͨͷ͔
    w %#..
    w -B[Z#.ͷϕʔεͱͳͬͨΞϧΰϦζϜ
    w -B[Z#."
    w %#.."NPSUJ[FE1JWPU'JMUFSJOH
    w -B[Z#."#
    w -B[Z#."#BMBODFE1SVOJOH
    w -B[Z#.
    w -B[Z#."#-PDBM5FSN1BSUJUJPOJOH
    -B[Z#.ͷ֤࠷దԽ͝ͱͷ଎౓
    w ֤࠷దԽͦΕͧΕͰ଎͘ͳ͍ͬͯΔ
    w ಛʹ#BMBODFE1SVOJOHͷߩݙ͕େ͖͍

    View Slide

  14. ײ૝
    w ༷ʑͳϥϯΩϯάϞσϧͰF⒏DJFODZΛแׅతʹධՁͨ͠ͷ͕ૉ੖Β͍͠
    w ΞϧΰϦζϜ͝ͱͷSPCVTUOFTT͕ݟΕͨ
    w ҰํͰɺ࣮ݧͷσʔληοτ͸$MVF8FC͔͠࢖͍ͬͯͳ͍
    w ੑ࣭͕ҟͳΔଞͷσʔληοτͰͷ࣮ݧ݁Ռ΋ݟ͍ͨ
    w -B[Z#.͸%PD*%PSJFOUFE#MPDL.BYΛ࢖͏%#..%#.8ͱಉ༷ɺϝϞϦ
    ফඅ͕ܹ͍͠
    w ϝϞϦ࢖༻ྔͱ଎౓ͱͷτϨʔυΦϑͷ؍఺΋΄͔ͬͨ͠

    View Slide

  15. &GGJDJFOUBOE&GGFDUJWF2VFSZ"VUP$PNQMFUJPO 4*(*3

    4JNPO(PH F#BZ*OD

    (JVMJP&SNBOOP1JCJSJ *45*$/3

    3PTTBOP7FOUVSJOJ 6OJWFSTJUZPG1JTB

    w F#BZͷ2"$ʹ͓͍࣮ͯ૷ͨ͠ΫΤϦϞʔυDPOKVODUJWFTFBSDIΛ঺հ
    w $POKVODUJWFTFBSDI͸QSFpYTFBSDIͱൺֱ͢Δͱ஗͍͕ɺF⒎FDUJWFOFTTͷ؍఺Ͱ༗༻
    w ΞϧΰϦζϜͱσʔλߏ଄Λ޻෉͠ɺDPOKVODUJWFTFBSDIΛߴ଎Խͨ͠

    View Slide

  16. ͭͷΫΤϦϞʔυ
    w 1SFpYTFBSDI
    w Ұൠతͳ2"$Ͱ࢖༻
    w ൃݟྗ͸ݶΒΕ͍ͯΔ
    w $POKVODUJWFTFBSDI
    w ඇৗʹߴίετ͕ͩ༗༻
    w ͜ΕΛߴ଎Խ͍ͨ͠
    ໰୊ઃఆ

    View Slide

  17. ࣮ݧ݁Ռ
    &GGFDUJWFOFTTͷධՁ
    w $POKVODUJWFTFBSDI͕QSFpYTFBSDI୯ମΑΓ΋ྑ͍݁ՌΛฦׂͨ͠߹
    w λʔϜҎ্ͷΫΤϦͰݦஶʹྑ͘ͳ͍ͬͯΔ
    ࠷ޙͷτʔΫϯͷ
    จࣈΛͲͷ͘Β͍
    ΫΤϦͱͯ͠࢖͏͔
    1SFpYTFBSDI୯ମ݅
    $POKVODUJWFTFBSDI୯ମ݅
    ߹ܭ݅

    View Slide

  18. $POKVODUJWFTFBSDIͷΞϧΰϦζϜ
    )FBQCBTFEΞϧΰϦζϜͱ'PSXBSECBTFEΞϧΰϦζϜ
    w DPNQMFUJPOσʔλߏ଄Λར༻ͯ͠ิ׬ީิΛநग़
    w 'PSXBSEJOEFYΛ࢖͏ํ๏ͱɺ'$ѹॖͨ͠σʔλߏ଄Λ࢖͏ํ๏
    ͕͋Δʢޙड़ʣ
    'PSXBSECBTFEΞϧΰϦζϜ
    )FBQCBTFEΞϧΰϦζϜ
    w σʔλߏ଄ͱͯ͠ώʔϓΛ࢖͏γϯϓϧͳํ๏

    View Slide

  19. $POKVODUJWFTFBSDIͰ࢖͏σʔλߏ଄
    σʔλߏ଄ͷ֓ཁͱνϡʔχϯά
    w ࣙॻ EJDUJPOBSZ

    w ֊૚ͷࣙॻΛ࢖͏
    w 'SPOU$PEJOH '$
    Ͱѹॖ͞ΕͨϒϩοΫ
    w $POKVODUJWFTFBSDIͰඞཁͳૢ࡞Λαϙʔτͭͭ͠ίϯύΫτˍߴ଎
    w ֤ϒϩοΫͷઌ಄ͷจࣈྻ͸ඇѹॖͰϔομͱͯ͠อଘ
    w ϒϩοΫΛ୳ͨ͢Ίʹ͸ɺ͜ͷϔομΛ෼୳ࡧ
    w ϒϩοΫαΠζΛ࣮ݧͰ࠷దԽ
    w αΠζͱ଎౓ͷτϨʔυΦϑͰ༏Ε͍ͯΔΛબ୒
    w ิ׬ީิू߹ DPNQMFUJPOT

    w ઀಄ࣙݕࡧ΍ɺ࠷ޙʹݕࡧ݁ՌΛฦ͢ͱ͖ʹ࢖༻
    w ݹయతʹ͸5SJF͕ଘࡏ
    w ࣙॻͱಉ༷ʹ'SPOU$PEJOH '$
    Ͱѹॖ͢Δํ๏͕͋Δ
    w ௕͍ΫΤϦͰ΋ߴ଎ͳ'$Λ࠾༻
    w సஔΠϯσοΫε JOWFSUFEJOEFY

    w ΫΤϦͷ֤λʔϜʹରͯ͠03ݕࡧ͢ΔͨΊʹඞཁ
    w ίϯύΫτ͔ͭߴ଎ͳ&MJBT'BOP &'
    Λ࢖͏
    w 3.2 3BOHF.JOJNVN2VFSJFT
    ؆ܿσʔλߏ଄
    w γϯάϧλʔϜΫΤϦͰͷݕࡧ͸͜ΕΛ࢖ͬͯߴ଎Խ
    ࣙॻͷόέοταΠζʹΑΔ࠷దԽ '$ͱ5SJFͷൺֱ
    సஔΠϯσοΫεͷѹॖΞϧΰϦζϜͷαΠζൺֱ

    View Slide

  20. ֤ख๏ͷ଎౓ൺֱ
    w )FBQCBTFE
    w )FBQ
    w 'PSXBSECBTFE
    w 'XEGPSXBSEJOEFYΛ࢖͏
    w '$'SPOU$PEJOHΛ࢖͏
    w #BTU8FCFS<>
    w )ZC )ZCJOEFY
    طଘख๏
    $POKVODUJWFTFBSDIͷ଎౓
    w 'PSXBSEϕʔεͷ'XEͱ'$͕ߴ଎
    w ΫΤϦ͕λʔϜͷ৔߹͸'XE͕଎͍
    w ͨͩ͠ฏۉ΄ͲϝϞϦ࢖༻ྔ͕ଟ͍
    w ͳ͓QSFpYTFBSDI͸ΫΤϦ͋ͨΓฏۉЖTFDҎԼ
    w λʔϜҎ্Ͱ͸ܻҧ͍ʹߴ଎ɾɾɾ

    View Slide

  21. ײ૝
    w $POKVODUJWFTFBSDI͸ΫΤϦϩά͕গͳ͍Ԡ༻ʹ͓͍ͯ༗༻ͦ͏
    w λʔϜͷॱং͕ೖΕସΘͬͨΫΤϦϩά͕༗ޮ׆༻Ͱ͖Δ
    w ߴ଎ԽͷͨΊͷ࣮૷͸͔ͳΓॏ͍
    w ༷ʑͳςΫ͍σʔλߏ଄ͱͦͷνϡʔχϯά͕ඞཁ
    w ؤுͬͯߴ଎Խͯ͠΋QSFpYTFBSDIͱ͸͔ͳΓ଎౓͕ࠩ͋Δ
    w λʔϜҎ֎ͷ৔߹ʹ͚ͩద༻͢ΔͳͲͷXPSLBSPVOE͕ߟ͑ΒΕΔ
    w ͋Δ͍͸QSFpYTFBSDIͰ͸े෼ͳ݅਺͕ฦͤͳ͍৔߹ʹ͚ͩద༻͢ΔͳͲ

    View Slide