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

IR Reading 2019秋 論文紹介 / IR Reading 2019Fall

Takuya Asano
October 27, 2019

IR Reading 2019秋 論文紹介 / IR Reading 2019Fall

IR Reading 2019秋 http://sigir.jp/post/irreading_2019fall/ で下記2本の論文を紹介しました。

1. Accelerated Query Processing Via Similarity Score Prediction (SIGIR 2019)
2. Fast Approximate Filtering of Search Results Sorted by Attribute (SIGIR 2019)

Takuya Asano

October 27, 2019
Tweet

More Decks by Takuya Asano

Other Decks in Research

Transcript

  1. *33FBEJOHळ࿦จ঺հ

    "DDFMFSBUFE2VFSZ1SPDFTTJOH7JB

    4JNJMBSJUZ4DPSF1SFEJDUJPO 4*(*3

    'BTU"QQSPYJNBUF'JMUFSJOHPG4FBSDI3FTVMUT
    4PSUFECZ"UUSJCVUF 4*(*3

    גࣜձࣾϝϧΧϦ

    ઙ໺୎໵

    View Slide

  2. "DDFMFSBUFE2VFSZ1SPDFTTJOH7JB
    4JNJMBSJUZ4DPSF1SFEJDUJPO 4*(*3

    w 63-
    w IUUQTEMBDNPSHDJUBUJPODGN JE
    w ݚڀͷ໨త
    w సஔΠϯσοΫεܕͷݕࡧΤϯδϯͰUPQLͷΫΤϦॲཧΛ

    ߴ଎Խ͢Δ͜ͱ
    w ํ๏
    w ΫΤϦॲཧதͷಈతͳจॻͷࢬמΓ QSVOJOH
    ʹண໨͠ɺ

    ػցֶशΛԉ༻ͯ͠ΞϧΰϦζϜΛվળ

    View Slide

  3. લఏ
    w ҰൠతͳసஔΠϯσοΫεΛ૝ఆ
    w λʔϜͱϙεςΟϯάϦετ͕อଘ͞Ε͍ͯΔ
    w ϙεςΟϯάϦετʹ͸จॻ*%ʢιʔτࡁΈʣͷू߹͕ೖ͍ͬͯΔ
    w λʔϜ͝ͱͷ࠷େείΞͳͲͷิॿతͳσʔλ΋֨ೲͰ͖Δͱ͢Δ
    w ͜ͷݚڀͰԾఆ͢ΔείΞؔ਺
    w جຊతʹ͸#.Λ૝ఆ
    w ΫΤϦʹؚ·ΕΔ֤λʔϜͷߩݙͷ଍͠߹Θͤ

    View Slide

  4. ΫΤϦॲཧ
    w EPDVNFOUBUBUJNF %""5
    ʹΑΔΫΤϦॲཧΛ૝ఆ
    w λʔϜ͝ͱʹϙεςΟϯάϦετΛॲཧ͢ΔͷͰ͸ͳ
    ͘ɺෳ਺ͷϙεςΟϯάϦετΛಉ࣌ʹΧʔιϧͰॲཧ
    w ͢΂ͯͷΧʔιϧΛҰ൪લʹηοτ͠ɺͦͷจॻͷΫΤ
    Ϧॲཧ͕ऴΘͬͨΒޙΖʹͣΒ͢
    w ͢΂ͯͷΧʔιϧ͕࠷ޙ·Ͱ͍ͬͨΒऴྃɺUPQLͷจ
    ॻू߹Λฦ͢

    View Slide

  5. ಈతࢬמΓʹΑΔΫΤϦॲཧ
    w ΫΤϦॲཧதʹɺͦΕ·Ͱʹॲཧͨ͠จॻͷUPQLͷείΞ Λܭࢉɾอ࣋͢Δʢ੨ઢʣ
    w ࠷ऴతʹ ʹऩଋ͢Δ
    w จॻͷείΞͷVQQFSCPVOEΛܭࢉ͢Δ
    w λʔϜ͝ͱͷ࠷େείΞ NBYTDPSF
    ΛΠϯσοΫεʹอଘ͓ͯ͘͠ͳͲ
    w VQQFSCPVOE͕ᮢ஋ ΑΓখ͍͞είΞͷจॻ͸҆શʹεΩοϓͰ͖Δ
    w ͜ͷํ๏͸ϩεϨεɺͭ·Γਖ਼֬ͳUPQLͷจॻ͕ಘΒΕΔ
    θ
    Θk
    θ

    View Slide

  6. طଘͷಈతࢬמΓΞϧΰϦζϜ
    w ޮ཰తͳΫΤϦॲཧͷํ๏ͱͯ͠.BY4DPSF 8"/%
    #.8 CMPDLNBY8"/%
    ͕͋Δ
    w ͢΂ͯجຊతͳઓུ͸લड़ͷಈతࢬמΓͱಉ͡
    w 8"/%͸ݡ͍ΞϧΰϦζϜͰ.BY4DPSFΑΓͨ͘͞Μ
    εΩοϓͰ͖ΔΑ͏ʹ͍ͯ͠Δ
    w #.8 CMPDLNBY8"/%
    ͸8"/%ΛϒϩοΫԽ͠ɺ
    ֤ϒϩοΫʹ౷ܭ৘ใΛ΋ͭΑ͏ʹ͞Βʹվྑͨ͠΋ͷ

    View Slide

  7. ͜ͷݚڀͷఏҊख๏
    w ͢΂ͯͷจॻʹ͍ͭͯΫΤϦॲཧ͢Δͱɺ ͸ɺUPQLείΞ ʹऩଋ͢Δ
    w ػցֶशΛ࢖ͬͯΫΤϦͷಛ௃ྔ͔Β࠷ऴతͳᮢ஋ Λ༧ଌ͢Δʢ੺ઢʣ
    w ͜ΕʹΑΓɺଟ͘ͷจॻΛεΩοϓͰ͖ΔΑ͏ʹͳΔߴ଎ԽͰ͖Δ
    θ Θk
    (D, Q)
    Θk
    (D, Q)

    View Slide

  8. 3FTFBSDI2VFTUJPOT
    w 32
    w సஔΠϯσοΫεΛॲཧ࢝͠ΊΔલʹ ΛਪఆͰ͖Δ͔ʁ
    w 32
    w ΋͠લ΋ͬͯ ΛਪఆͰ͖ΔͳΒɺͦͷࣄલ஌ࣝʹΑͬͯ

    Ͳͷఔ౓ΫΤϦॲཧΛ଎͘Ͱ͖Δ͔ʁ
    w 32
    w ͷࣄલ஌ࣝΛ࢖ͬͯ.BY4DPSFɺ8"/%ɺ#.8ͷΑ͏ͳख๏
    ΛվྑͰ͖Δ͔ʁ
    Θk
    (D, Q)
    Θk
    (D, Q)
    Θk
    (D, Q)

    View Slide

  9. ΦϥΫϧʢਆୗʣΛ࢖࣮ͬͨݧ
    w 32΋͠લ΋ͬͯ ΛਪఆͰ͖ΔͳΒɺͦͷࣄલ஌ࣝʹΑͬͯͲͷఔ
    ౓ΫΤϦॲཧΛ଎͘Ͱ͖Δ͔ʁ΁ͷճ౴
    w ΛΫΤϦॲཧͷલʹܭࢉ͓͖ͯ͠ɺ֤ࢬמΓΞϧΰϦζϜʹᮢ஋ͱ͠
    ͯ౉ͯ͠͠·͏
    w ࠷దͳᮢ஋͕෼͔͍ͬͯΔঢ়ଶͰͷɺ."9ύϑΥʔϚϯεΛܭଌ
    w ࣮ݧ৚݅
    w (PWσʔληοτɺ ΫΤϦͰ࣮ߦ࣌ؒΛϛϦඵ୯ҐͰܭଌ
    w ࢬמΓΞϧΰϦζϜͱͯ͠.BY4DPSF 8"/% #.8ͷͭΛ࢖ͬͯൺֱ
    Θk
    (D, Q)
    Θk
    (D, Q)

    View Slide

  10. ΦϥΫϧʢਆୗʣΛ࢖࣮ͬͨݧ
    w ΦϥΫϧΛ࢖͏ͱϕʔεϥΠϯ
    ΑΓ΋ʙ΄Ͳߴ଎Խ
    w Λਖ਼֬ʹਪఆ͢Δ͜ͱͰɺ
    ߴ଎Խͷ༨஍͕͋Δʂ

    Θk
    (D, Q)

    View Slide

  11. ֬཰Ϟσϧϕʔεͷᮢ஋༧ଌ
    w ֬཰ϞσϧΛ࢖ͬͨ༧ଌ
    w ༧ଌ࣌ؒࠐΈͰશମͷΫΤϦॲཧΛ଎͘͠ͳ͍ͱ͍͚ͳ͍ͷͰɺγϯϓϧͳϞσϧͰ
    w ͷ෼෍͕Ψ΢ε෼෍Ͱ͋ΔͱԾఆ͠ɺύϥϝʔλΛֶशͯ͠༧ଌ෼෍ͱ͢Δ
    w ͦͷ෼෍ͷ্Ͱ࠷దͳ Λਪఆ͢Δʢ ʣ
    Θk
    ̂
    Θk
    q = ?

    View Slide

  12. w PWFSFTUJNBUFͨ͠ͱ͖ʹΑΓϖφϧςΟΛ༩͍͑ͨ
    w ֬཰Ϟσϧʹଛࣦؔ਺Λ૊Έ߹ΘͤΔ
    w ࠓճ͸PWFSFTUJNBUFͨ͠ͱ͖ͷଛࣦΛ ഒͱ
    ͢ΔΑ͏ͳઢܗͷଛࣦؔ਺ʢ੨ઢʣΛߟ͑Δ
    α
    ֬཰Ϟσϧϕʔεͷᮢ஋༧ଌ
    w .JOJNVN#BZFT3JTL .#3

    w ༧ଌ෼෍ͷ্Ͱɺଛࣦͷظ଴஋͕࠷খͱ
    ͳΔΑ͏ͳ Λબ୒Ͱ͖Δ
    w ࠷దͳ ͸ ͷRVBOUJMFͱͳΔ
    ͜ͱ͕ূ໌͞Ε͍ͯΔ<>
    w ͨͱ͑͹ɺ ͷͱ͖RVBOUJMF
    ࢛෼Ґ఺
    ͕࠷ద
    w Ϟσϧʹ͸#BZFTJBO-JOFBS3FHSFTTJPO
    #-3
    Λ࢖ͬͨ
    w ղੳతʹʢดܗࣜͰʣղ͕ٻΊΒΕΔͷ
    Ͱߴ଎ʹܭࢉͰ͖Δ
    ̂
    Θk
    ̂
    Θk
    1/(α + 1)
    α = 3
    <>1'$ISJTUP⒎FSTFOBOE'9%JFCPME0QUJNBMQSFEJDUJPOVOEFSBTZNNFUSJDMPTT&DPOPNFUSJD5IFPSZ
    r

    View Slide

  13. ଛࣦϕʔεͷᮢ஋༧ଌ
    w ଛࣦؔ਺Λ࢖ͬͨػցֶशϞσϧ΋ߟ͑Δ
    w ༩͑ΒΕͨଛࣦؔ਺ʹରͯ͠௚઀ֶश
    w ଛࣦؔ਺ʹ͸ඇରশʹͨ͠)VCFSMPTTΛ࠾༻
    w ඍ෼Մೳ
    w ϚϧνϨΠϠʔύʔηϓτϩϯ .-1

    w .-1-

    ૚ʢӅΕ૚͕ʣͷ.-1
    w .-1-

    ૚ʢӅΕ૚͕ʣͷ.-1
    w ͍ͣΕ΋$16Ͱޮ཰తʹܭࢉՄೳ

    View Slide

  14. Ϟσϧͷֶश
    w Ϟσϧͷֶशʹ࢖ͬͨΫΤϦϩά
    w .JDSPTPGU2VFSZMPH
    w ΫΤϦ
    w σʔληοτ
    w (PWສ݅
    w $MVF8FC#ສ݅
    w ֶशࡁΈϞσϧͷ༧ଌ଎౓
    w NT .-1-

    w NT .-1-

    w NT #-3

    w ͍ͣΕ΋NTҎ಺ͳͷͰे෼଎͍
    ֶशʹ࢖ͬͨಛ௃ྔ

    View Slide

  15. ਪఆͨ͠ᮢ஋Λ࢖ͬͨ

    ΫΤϦॲཧ
    PWFSFTUJNBUFͯͨ͠Β࠶ܭࢉ
    ώʔϓαΠζ͕༰ྔʢLݸʣΛ௒͑ͨͱ͖ͷॲཧ
    ώʔϓͷ࠷খείΞΑΓେ͖͚Ε͹ώʔϓʹೖΕΔ
    είΞͷVQQFSCPVOE͕ Λ௒͑Δ

    จॻ͚ͩείΞܭࢉ
    θ
    ᮢ஋ͷ༧ଌ஋ Λ ͷॳظ஋ͱ͢Δ
    ̂
    Θk
    θ

    View Slide

  16. "MHPSJUINʹΑΔ

    ΫΤϦॲཧͷ࣮ߦ଎౓
    ϕʔε

    ϥΠϯ
    ϕʔε

    ϥΠϯ
    ύʔηϓτϩϯ #BZFTJBO-JOFBS3FHSFTTJPO
    গ͚ͩ͠଎͘ͳͬͨ

    View Slide

  17. "MHPSJUIN͕஗͍ཧ༝
    w PWFSFTUJNBUFͨ͠ͱ͖ʹ
    ࠶ܭࢉ͕૸Δ
    w ͜Ε͕ॏ͍

    View Slide

  18. "MHPSJUINʢվྑ൛ʣ
    w ໰୊ʹͳ͍ͬͯΔλʔϜͷ૊Έ߹Θͤ
    ʢαϒΫΤϦʣΛܭࢉ
    w αϒΫΤϦʹϚον͢Δจॻ͚ͩ

    ݟΕ͹Α͍
    w จॻ͕͔ͳΓߜΕΔͷͰߴ଎ʹ

    ࠶ܭࢉՄೳ

    View Slide

  19. "MHPSJUINʹΑΔ

    ΫΤϦॲཧͷ࣮ߦ଎౓
    w தԝ஋Ͱ࠷େͰ଎͘ͳͬͨ
    w ͨͩ͠ಉ͡ Ͱൺ΂ͨͱ͖
    w ࠷ద஋ͱൺ΂Δͱ͘Β͍
    w "MHPSJUIN͸RVBOUJMF͕େ͖͍

    ʢΞάϨογϒͳʣ΄͏͕଎͍
    q = 0.5

    View Slide

  20. 3FTFBSDI2VFTUJPOT
    w 32
    w సஔΠϯσοΫεΛॲཧ࢝͠ΊΔલʹ ΛਪఆͰ͖Δ͔ʁ
    w ˠώʔϓͷᮢ஋͕༧ଌՄೳͰ͋Δ͜ͱΛࣔͨ͠
    w 32
    w ΋͠લ΋ͬͯ ΛਪఆͰ͖ΔͳΒɺͦͷࣄલ஌ࣝʹΑͬͯ

    Ͳͷఔ౓ΫΤϦॲཧΛ଎͘Ͱ͖Δ͔ʁ
    w ˠಈతࢬמΓΞϧΰϦζϜͷ଎౓͕Ҏ্վળ͢Δ͜ͱΛࣔͨ͠
    w 32
    w ͷࣄલ஌ࣝΛ࢖ͬͯ.BY4DPSFɺ8"/%ɺ#.8ͷΑ͏ͳख๏ΛվྑͰ͖Δ͔ʁ
    w ˠ͢΂ͯͷख๏ͰΞϧΰϦζϜతͳվળ͕ՄೳͰ͋Δ͜ͱΛࣔͨ͠
    Θk
    (D, Q)
    Θk
    (D, Q)
    Θk
    (D, Q)

    View Slide

  21. ײ૝
    w ܾఆతͳํ๏ʢ Λ࢖͏ϕʔεϥΠϯʣͰ΋े෼Ͱ͸ʁ
    w 5BCMFΛΈΔͱɺϕʔεϥΠϯ͔Βͷࠩ͸ͦΕ΄Ͳେ͖͘ͳ͍
    w "MHPSJUINͰଟগվળ͢Δͱ͸͍͑ɺࠩ෼͸͍͍ͤͥఔ౓
    w ࢬמΓΞϧΰϦζϜͷࠩͷ΄͏͕େ͖͍
    w ࢓૊Έ͕γϯϓϧͳ ͷ΄͏͕औΓճ͠΍͍͢ʢػցֶश͍Βͳ͍͠ʣ
    w ΋ͬͱ͍͍ϞσϧΛ࡞ΕΔ༨஍͸͋Δ
    w ಛ௃ྔ΍ػցֶशΞϧΰϦζϜͷվળ͕Մೳ
    w ਅͷᮢ஋ʢΦϥΫϧʣʹ͚ۙͮ͹ͦΕͳΓʹ଎͘ͳΔ
    Qk
    Qk

    View Slide

  22. 'BTU"QQSPYJNBUF'JMUFSJOHPG4FBSDI
    3FTVMUT4PSUFECZ"UUSJCVUF 4*(*3

    w 63-
    w IUUQTEMBDNPSHDJUBUJPODGN JE
    w ໰୊ઃఆ
    w ଐੑʹΑΔιʔτॱΛकΓͳ͕Β%$(ͷΑ͏ͳϝτϦοΫΛ࠷దԽ͢Δ໰୊
    w ͜ͷݚڀͷ੒Ռ
    w ߴ଎ʹ΄΅࠷దͳ݁ՌΛฦ͢ΞϧΰϦζϜΛఏҊ
    w ۙࣅʹΑΔΤϥʔ཰ͱ࠷ѱ࣌ؒܭࢉྔͷཧ࿦อূ͋Γ
    w ࣮ݧతʹ΋ߴ଎ੑͱΤϥʔ཰ͷόϥϯε͕Α͍͜ͱ͕ࣔ͞Εͨ

    View Slide

  23. ໰୊ઃఆ'JMUFSJOH!L
    w ௨ৗɺϢʔβ͕ݕࡧͰཉ͍͠ͷ͸
    SFMFWBODFΛ࠷େԽͨ͠4&31
    w 8FCݕࡧɺ4/4ɺετϦʔϛϯά
    αʔϏεɺ&$αΠτͳͲ
    w ଟ͘ͷαʔϏεͰ͸ɺଐੑʢ࣌ؒ΍Ձ֨
    ͳͲʣʹΑΔιʔτʹରԠ͍ͯ͠Δ
    w ͨͱ͑͹"NB[POͰJQIPOFͷݕࡧ݁
    ՌΛՁ֨ͷ͍҆ॱʹ͢ΔͱSFMFWBODFͷ

    ௿͍ΞΠςϜ্͕ʹདྷΔͨΊɺϢʔβମ
    ݧΛଛͳ͏
    BNB[PODPKQͰՁ֨ͷ͍҆ॱͰJQIPOFΛݕࡧ

    ʢՈిˍΧϝϥΧςΰϦʣ

    View Slide

  24. ໰୊ઃఆ'JMUFSJOH!L
    w 'JMUFSJOH!L
    w SFMFWBODFͷϦετ ʢॱং͖ͭʣ͕

    ༩͑ΒΕΔ
    w ϝτϦοΫ2Λ࠷େԽ͢Δɺ࠷େL݅ͷ෦෼Ϧετ

    Λฦ͢
    R = < r1
    , . . . , rn
    >

    View Slide

  25. 0QUJNBM'JMUFSJOH 015

    w ࣌ؒܭࢉྔ Ͱ࠷దͳ݁Ռʢۙࣅͳ͠ʣΛܭࢉͰ͖Δ
    w ಈతܭը๏ʢϝϞԽʣΛ࢖ͬͯܭࢉ
    w O L͕େ͖͘ͳΔͱ஗͍
    Θ(nk)

    View Slide

  26. ώϡʔϦεςΟΫεʹΑΔํ๏
    w $VUP⒎015
    w ͋Β͔͡ΊܾΊͨSFMFWBODFͷᮢ஋ͰϑΟϧλ͔ͯ͠Β015Ͱ͞ΒʹϑΟϧλ
    w ࠷ѱέʔεͰ͸Τϥʔ཰ʢਫ਼౓ʣͷอূ͕ͳ͍
    w ᮢ஋ͷઃఆʹΑΔͨΊ
    w ࣌ؒܭࢉྔ͸վળͤͣ࠷ѱέʔεͰ
    w 5PQL015
    w UPQLͷϦετΛٻΊɺͦΕΛ͞Βʹ015ͰϑΟϧλ͢Δ
    w LݸະຬͷϦετʹ͢Δ͜ͱͰϝτϦοΫΛ࠷େԽͰ͖Δ͜ͱ΋͋ΔͨΊ
    w PQUJNBMʢۙࣅʹΑΔΤϥʔ཰͕ʣͰ͋Δ͜ͱ͕ূ໌Ͱ͖Δ
    w ࣌ؒܭࢉྔ͸࠷ѱͰ
    Θ(nk)
    Θ(n log k + k2)

    View Slide

  27. Џ'JMUFSJOH
    w $VUP⒎΍5PQLͳͲͷώϡʔϦεςΟΫεͰ͸Τϥʔ཰ΛίϯτϩʔϧͰ͖ͳ͍
    w Џ'JMUFSJOH͸Τϥʔ཰ΛЏҎԼʹอূͨۙ͠ࣅղΛɺ

    ࣌ؒܭࢉྔ Ͱݟ͚ͭΔʢ ʣ
    w Џ'JMUFSJOH͸ҎԼͷεςοϓ͔ΒͳΔ
    w 3JHIUQSVOJOH
    w %JTDSFUJ[BUJPO
    w 5ISFTIPMEJOH
    w ֤ϑΣʔζͰى͜Γ͏Δ࠷ѱέʔεʹɺޙͷϑΣʔζͰରԠ͢Δ͜ͱͰ

    ܭࢉ࣌ؒΛ࡟͍ͬͯ͘
    Θ(n + k2 log(1−ϵ)
    (ϵ/k)) 0 < ϵ < 1

    View Slide

  28. ࣮ݧ݁Ռ
    w Џ'JMUFSJOH͸5PQL015ͱݞΛฒ΂Δ଎͞Ͱɺখ͍͞Τϥʔ཰Λୡ੒͍ͯ͠Δ
    w Џ'JMUFSJOHͷ ɺ Ͱ͸Τϥʔ཰Λ΄΅ʹ཈͑ͳ͕Β015ΑΓ΋அવ଎͍
    ϵ = 0.01 ϵ = 0.001

    View Slide