$30 off During Our Annual Pro Sale. View Details »

CommonLitコンペで学んだこと

nogawanogawa
October 12, 2023
1.1k

 CommonLitコンペで学んだこと

nogawanogawa

October 12, 2023
Tweet

Transcript

  1. ©2023 Wantedly, Inc.
    CommonLitίϯϖͰֶΜͩ͜ͱ
    Custom HeaderʹΑΔTransformerੑೳվળ
    ΈΜͳͷPythonษڧձ #98


    Oct. 12 2023 - @nogawanogawa

    View Slide

  2. ©2023 Wantedly, Inc.
    CommonLit ίϯϖ ?
    CommonLit - Evaluate Student Summaries

    View Slide

  3. ©2023 Wantedly, Inc.
    CommonLit - Evaluate Student Summaries
    KaggleʢσʔλαΠΤϯείϯϖςΟγϣϯϓϥοτ
    ϑΥʔϜʣͰ։࠵͞Ε͍ͯͨେձ


    • ظؒ: 7/13 ~ 10/12 (3͔݄ɺࠓேऴྃʂ)


    • ࢀՃऀ: 2106νʔϜ
    • ࣗ෼ͷ࠷ऴॱҐ͸25ҐͰͨ͠

    View Slide

  4. ©2023 Wantedly, Inc.
    CommonLit - Evaluate Student Summaries
    • খֶߍ3೥ੜʙߴߍ3೥ੜͷੜెʹΑΔจষͷཁ໿͕ςʔϚ


    • ੜె͕࡞੒ͨ͠ཁ໿จͷ࣭Λهड़಺༰(contentʣɺޠኮɾจ๏ʢwordingʣͷ2ͭͷ؍఺Ͱ࠾఺


    ◦ ࣮ࡍʹਓ͕ؒ࠾఺ͨ݁͠Ռʹ࠷΋͍ۙػցֶशϞσϧΛ࡞ͬͨਓ͕উར
    จষΛཁ໿
    ੜె
    ઌੜ
    XPSEJOH఺
    DPOUFOU఺
    XPSEJOH఺
    DPOUFOU఺
    ػցֶशϞσϧ
    ͳΔ΂͘ਓؒͷ


    ࠾఺݁Ռʹ͍ۙ


    ϞσϧΛ࡞Γ͍ͨʂ
    ࢀՃऀ͸͜ΕΛ࡞͍ͬͯͨ
    ࠾఺
    ༧ଌ

    View Slide

  5. ©2023 Wantedly, Inc.
    TransformerͱCustom Header

    View Slide

  6. ©2023 Wantedly, Inc.
    TransformerΛ࢖ͬͨجຊతͳղ๏
    • ࣗવจΛ࢖͏ίϯϖͰ͸Α͘Transformer͕࢖༻͞ΕΔ


    • χϡʔϥϧωοτϫʔΫͷݱࡏओྲྀͷΞʔΩςΫνϟ


    • ߏ଄͸ӈਤͷΑ͏ʹͳ͍ͬͯΔ


    ◦ ChatGPT΋ཪଆ΋͜Εͷ஥ؒ


    • ࢖͏͜ͱࣗମ͸ҙ֎ͱ؆୯


    ◦ HuggingFaceͷTransformersϥΠϒϥϦΛ࢖͑͹਺
    ेߦఔ౓ॻ͚͹खܰʹ࢖͑Δ


    ◦ ੈքதͷਓֶ͕शࡁΈϞσϧͷσʔλΛެ։ͯ͘͠Εͯ
    ͍ͯɺDLͯ͠࢖͏͚ͩ

    View Slide

  7. ©2023 Wantedly, Inc.
    Transformers
    ࢖༻͍ͨ͠ϞσϧͷछྨΛબ୒
    ֶशࡁΈϞσϧΛμ΢ϯϩʔυ
    ೉͘͠ͳ͍ʂ
    ۩ମతʹղ͖͍ͨ໰୊ͷͨΊʹඍௐ੔(fine tune)
    ͢ΔίʔυΛॻ͍ͨͱͯ͠΋ϓϥε਺ेߦఔ౓
    • ࢖༻͢Δ͚ͩͳΒ਺ߦͰॻ͚Δ

    View Slide

  8. ©2023 Wantedly, Inc.
    ਫ਼౓޲্ͷํ਑
    • લॲཧ


    • ଟ͘ͷσʔλ͸͋·Γ͖Ε͍Ͱ͸ͳ͍ʢϊΠζͱͳΔ΋ͷ͕ଟ͍ʣࣄ͕ଟ͘ɺͦΕΒΛআڈ͢Δ


    • ࣄલֶशࡁΈϞσϧͷมߋ


    • ࢖͑ΔֶशࡁΈϞσϧʹ΋ͨ͘͞Μछྨ͕͋ΔʢBertɺRobertaɺDebertaɺetc… ʣ


    • ֶशσʔλͷ૿ڧ


    • ΑΓଟ͘ɾଟ༷ͳσʔλͰֶशͨ͠Ϟσϧ͕ڧ͘ͳΓ͕ͪ


    • Ξϯαϯϒϧ


    • ෳ਺ͷػցֶशϞσϧΛ࢖ͬͨଟ਺ܾ


    • Transformerͷߏ଄ࣗମʹςίೖΕ


    • Custom Header ʢPoolerʣ


    • etc…
    ࠓ೔͸͜͜Λ࿩͍ͨ͠ʂ
    Transformers͸؆୯ʹ࢖͑Δ & ߴ͍ਫ਼౓ͷϞσϧΛ࢖༻Ͱ͖Δɻͨͩ͠ɺKaggleͳͲͷίϯϖͰ͸͜Ε
    Λ࢖͏ͷ͕ελʔτϥΠϯͰɺίϯϖͰॱҐΛ্͛Δʹ͸͔͜͜Βߋʹ޻෉͕ඞཁ

    View Slide

  9. ©2023 Wantedly, Inc.
    Custom Header (Pooler)ʹΑΔਫ਼౓޲্ࡦ
    • σϑΥϧτͰTransformerΛ࢖༻͢Δͱɺೖྗͷ࠷ॳͷτʔΫϯ (CLSτʔΫϯ)ͷ࠷ऴ૚ग़ྗΛ΋ͱʹ༧ଌ
    ࠷ॳͷτʔΫϯग़ྗ͚ͩΛ࢖༻ɺଞͷτʔΫϯͷग़ྗ͸ࣺ͍ͯͯΔ
    →௚ײతʹ͸ඇৗʹ΋͍ͬͨͳ͘ײ͡Δ


    ʢจষશମͷग़ྗΛ࢖ͬͨ΄͏͕දݱྗ͸ߴ͘ͳΓͦ͏ʹࢥ͑Δʣ

    View Slide

  10. ©2023 Wantedly, Inc.
    Custom Header (Pooler)ʹΑΔਫ਼౓޲্ࡦ
    • Custom Header (Pooler) : จશମͷग़ྗΛ࢖ͬͯpoolingΛߦ͍ɺ࠷ऴ૚ʹೖྗ


    • poolingͷྫ: ฏۉɺ࠷େɺ࠷ऴ4૚ͷग़ྗΛͭͳ͛Δɺetc…
    ࢀߟ: https://www.kaggle.com/code/rhtsingh/utilizing-transformer-representations-efficiently
    ೉ͦ͠͏ʹݟ͑ͯ


    ҙ֎ͱ਺ߦͰॻ͚Δ
    ྫ: Average pooling

    View Slide

  11. ©2023 Wantedly, Inc.
    ΦϦδφϧCustom HeaderʹΑΔਫ਼౓޲্ࡦ
    • ࠓճࣗ෼͕΍ͬͯҰ൪͏·͘ߦͬͨํ๏


    • ཁ໿ର৅ͷจͷՕॴ͚ͩΛaverage
    poolingͯ͠࢖༻


    • ղऍ


    • ࠓճ͸ੜెͷཁ໿จͷ࣭ΛධՁ͢Δͷ͕໨త


    • ධՁର৅ͷ෦෼͚ͩ࢖༻͢Δ͜ͱͰ࠷ऴ૚
    ΁ͷϊΠζΛݮΒ͢

    View Slide

  12. ©2023 Wantedly, Inc.
    ·ͱΊɿCommonLitίϯϖͰֶΜͩ͜ͱ
    •TransformerϞσϧ͸ͦ͜·Ͱۤ࿑ͤͣ࢖͑ͯߴੑೳɻͨͩ͠Ԟ΋ਂ͍ɻ


    • TransformersΛ࢖͏ͱઌਓ͕࡞ͬͨߴੑೳͷֶशࡁΈϞσϧΛ࢖༻Ͱ͖Δ


    ◦ ࢖͏͚ͩͳΒͦ͜·Ͱ೉͘͠ͳ͍


    • ͨͩ͠࠷ۙͷKaggleͳͲͷNLPίϯϖͰ͸ɺTransformerΛ࢖͏ͷ͸͋͘·ͰελʔτϥΠϯ


    ◦ σʔλαΠΤϯςΟετతʹ͸΋͏Ұา౿ΈࠐΜͰ͍͖͍ͨ


    • TransformerͷੑೳΛߋʹ޲্ͤ͞ΔΞϓϩʔνͷ1ͭ: Custom Header (Pooler)


    ◦ Transformer͕λεΫΛղ͖΍͍͢Α͏ʹ࠷ޙͷPoolingΛௐ੔͢Δͱ্ख͍͘͘͜ͱ͕͋Δ

    View Slide