Finetuned Language Models Are Zero-Shot Learners（最先端NLP2022）

論文URL:
https://openreview.net/forum?id=gEZrGCozdqR

1. ࿦จ঺հ: Finetuned Language Models Are Zero-Shot Learners Jason Wei, Maarten Bosma, Vincent Y. Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, Quoc V. Le Google Research ಡΈखɿүপ େ @ICLR2022 ஫ه͕ͳ͍ݶΓɺਤද͸ຊ࿦จ͔ΒͷҾ༻
2. • എܠ֓ཁ • ఏҊख๏σʔληοτ • ධՁ࣮ݧ • ߟ࡯ ໨࣍ 2
3. • λεΫͷࢦࣔ΍ྫࣔΛ༩͑Δ͜ͱͰɺ೚ҙͷλεΫΛղ͘͜ͱ͕Ͱ͖ΔݴޠϞσϧ͕ొ৔ • ൚༻ݴޠϞσϧ࣮ݱ΁ͷେ͖ͳҰาͱͯ͠஫໨ΛཋͼΔ GPT-3ͷొ৔ 3 https://beta.openai.com/playground 1BSBQISBTF
   UIF
   GPMMPXJOH
   UFYU
   &YBNQMFT *
   MPWF
   GSVJUT
   
   *
   FOKPZ
   FBUJOH
   GPPE
   *
   IBWF
   MVODI
    *
   FBU
   MVODI ೖྗ ग़ྗ 5SBOTMBUF
   UIF
   GPMMPXJOH
   UFYU
   JOUP
   +BQBOFTF
   )PX
   BSF
   ZPVʁ ͓ݩؾͰ͔͢ʁ λεΫͷࢦࣔ λεΫͷྫࣔ (15
4. • (15͸8FC্ͷ௒େྔͷจষʢ ԯ୯ޠʣΛಡΈࠐΉʢ࣍ͷ୯ޠΛ༧ଌ͢Δʣ͜ͱͰֶश – ͦͷதʹ͸Լهͷ2"αΠτͳͲ΋ؚ·Εɺ༷ʑͳ࣭໰ͱͦΕʹର͢Δճ౴Λ҉ʹֶश͍ͯ͠ΔͷͰ͸ͳ͍͔ → λεΫͷࢦࣔͱͦΕʹର͢Δճ౴Λ໌ࣔతʹֶशͤ͞Ε͹ɺ(15ͷΑ͏ͳܳ౰͕Ͱ͖ͳ͍ͩΖ͏͔ – (15ΑΓ΋ খ͍͞ϞσϧͰ ߴ͍҆ఆͨ͠ੑೳΛ࣮ݱͰ͖ͳ͍ͩΖ͏͔ ͳͥGPT-3͸ͦͷΑ͏ͳܳ౰͕Ͱ͖Δͷ͔ 4 ग़య: https://c4-search.apps.allenai.org/
5. • JOTUSVDUJPO
   UVOJOH
   λεΫͷࢦࣔ΍ྫࣔΛϓϩϯϓτͱͯ͠Ճ͑ͨϚϧνλεΫֶश • ֶश࣌ʹ͸༩͍͑ͯͳ͍λεΫͱͦͷࢦࣔΛ༻͍ͯධՁ͢Δ͜ͱͰɺ Ϟσϧ͕ະ஌ͷࢦࣔͱλεΫʹ൚ԽͰ͖Δ͔ධՁ λεΫͷࢦࣔͱͦΕʹର͢Δճ౴Λֶश͢ΔJOTUSVDUJPO
   UVOJOHΛఏҊ 5 5SBOTMBUF
   UIF
   GPMMPXJOH
   UFYU
   JOUP
   +BQBOFTF )PX
   BSF
   ZPV 8IBU
   JT
   UIF
   TFOUJNFOU
   PG
   UIJT
   NPWJF
   SFWJFX *
   MJLF
   TQPSUT 5IJT
   NPWJF
   JT
   GVO *
   EPO`U
   MJLF
   UIJT
   GJMN ͓ݩؾͰ͔͢ʁ ࢲ͸ӡಈ͕޷͖Ͱ͢ɻ QPTJUJWF OFHBUJWF ֶश ධՁ ௨ৗͷϚϧνλεΫֶश 1BSBQISBTF
   UIF
   GPMMPXJOH
   UFYU *
   MPWF
   GSVJUT *
   HP
   UP
   TDIPPM *
   FOKPZ
   FBUJOH
   GSVJU *
   BUUFOE
   TDIPPM ຋༁ ײ৘෼ੳ ݴ͍׵͑ + + + + + + ⋮ ⋮ ⋮ ਺ेݸͷ λεΫ ֶश͍ͯ͠ ͳ͍λεΫ ϓϩϯϓτ Ϟσϧ
6. ۙ೥ྨࣅͷऔΓ૊Έ͕૬͙࣍ 6 '-"/
   ঺հ࿦จ 5<> /BUVSBM
   *OTUSVDUJPOT W<> W<> .FUB*$-<> த৺૊৫ ൃදձٞ ϓϩϯϓτ ͷ಺༰ #JH4DJFODF XPSLTIPQ *$-3 *$-3 W
   "$- W
   BS9JW ౤ߘத /""$- ࢦࣔ ˎྫࣔΛՃ࣮͑ͨݧ༗Γ ࢦࣔ ࢦࣔ ྫࣔ ˎෛྫ౳Ճ࣮͑ͨݧ΋༗Δ͕ɺ λεΫͷࢦࣔ ྫ͕ࣔ࠷ߴੑೳ ྫࣔ ˎࢦࣔΛՃ࣮͑ͨݧ༗Γ Ϟσϧ -B.%"15
   EFDPEFSPOMZ
   # 5-. FODPEFSEFDPEFS
   ## W
   #"35 FODPEFSEFDPEFS
   . W
   5-. (15 EFDPEFSPOMZ
   . [1] Sanh et al., Multitask Prompted Training Enables Zero-Shot Task Generalization. ICLR, 2022. [2] Mishra et al., Cross-Task Generalization via Natural Language Crowdsourcing Instructions. ACL, 2022. [3] Wang et al., Benchmarking Generalization via In-Context Instructions on 1,600+ Language Tasks. arXiv, 2022. [4] Min et al., MetaICL: Learning to Learn In Context. NAACL, 2022. (JUIVCͷϨϙδτϦ΍ϓϩϯϓτ ऩूαΠτͷυϝΠϯΛ΋ͱʹಉఆ
7. • ैདྷͷσʔληοτʹՃ͑ɺϓϩϯϓτΛऩू͢ΔྲྀΕ͕࢝·͍ͬͯΔ ϓϩϯϓτΛऩू͢ΔϓϥοτϑΥʔϜ͕ొ৔ 7 [1] Bach et al., PromptSource: An Integrated Development Environment and Repository for Natural Language Prompts. ACL demo, 2022. https://github.com/bigscience-workshop/promptsource [2] Wang et al., Benchmarking Generalization via In-Context Instructions on 1,600+ Language Tasks. arXiv, 2022. https://instructions.apps.allenai.org/ 1SPNQUTPVSDFʢ5ͷֶशධՁʹ࢖༻ʣ<> /BUVSBM
   *OTUSVDUJPOT<>
8. • എܠ֓ཁ • ఏҊख๏σʔληοτ • ධՁ࣮ݧ • ߟ࡯ ໨࣍ 8
9. • طଘͷσʔληοτΛ༻ҙ͠ɺ֤σʔληοτΛΫϥελʹ෼ྨ • Ұ෦ͷΫϥελͰֶशΛߦ͍ɺֶशʹ࢖༻͠ͳ͍ΫϥελͰධՁ͢Δ͜ͱͰݫີʹ[FSPTIPUੑೳΛଌΔ σʔληοτͷऩूɾ෼ྨ 9 /BUVSBM
   -BOHVBHF
   6OEFSTUBOEJOH
   ෼ྨλεΫ /BUVSBM
   -BOHVBHF
   (FOFSBUJPO
   ੜ੒λεΫ * ߋʹɺ3FBE
   DPNQ
   X
   DPNNPOTFOTFΫϥελͰධՁ͢Δࡍ͸3FBEJOH
   DPNQͱ$PNNPOTFOTFΫϥελ͸ֶशʹ༻͍ͣɺ /BUVSBM
   MBOHVBHF
   JOGFSFODFΫϥελͰධՁ͢Δࡍ͸1BSBQISBTFΫϥελ͸ֶशʹ༻͍ͳ͍ɻٯ΋ಉ༷ɻ
10. • ֤σʔληοτʹ͍ͭͯछྨͷςϯϓϨʔτʢϓϩϯϓτʣΛ࡞੒ – ྫʣ/BUVSBM
    -BOHVBHF
    *OGFSFODFʢؚҙؔ܎ೝࣝʣͷ৔߹ • λεΫͷଟ༷ੑΛ֬อ͢ΔͨΊʹɺσʔληοτͷೖྗͱग़ྗΛೖΕସ͑ͨςϯϓϨʔτ΋࡞੒ – ྫʣײ৘ۃੑ෼ྨσʔληοτΛ༻͍ͯɺײ৘ۃੑΛೖྗʹ঎඼ϨϏϡʔΛੜ੒͢ΔςϯϓϨʔτΛ࡞੒ ςϯϓϨʔτͷ࡞੒ 10
11. • ࡞੒͞ΕͨςϯϓϨʔτʹσʔληοτதͷ֤σʔλΛ౰ͯ͸ΊݴޠϞσϧʹೖྗ͠ɺճ౴Λग़ྗɻ ਖ਼͍͠ճ౴ͷ໬౓Λ࠷େԽ͢ΔΑ͏ʹֶश • ෼ྨੜ੒λεΫڞʹɺੜ੒͞Εͨճ౴ͱਖ਼͍͠ճ౴Λൺֱͯ͠ධՁ – ͔͠͠෼ྨλεΫͰ͸ਖ਼͍͠ճ౴ͷදݱ͸༷ʑʢྫ͑͹্هྫͰ͸ɺlZFTzҎ֎ʹlFOUBJMNFOUz͕༗ΓಘΔʣ – ͦ͜ͰςϯϓϨʔτʹl0QUJPOT
    ZFT
    OPzΛՃ͑Δ͜ͱͰɺճ౴͸lZFTOPzͷ͍ͣΕ͔Ͱ͋Δ͜ͱΛࢦࣔ Ϟσϧͷֶश/ධՁ 11 3FBE UIF GPMMP XJOH ݴޠϞσϧ ʜ ZFT OP ZFT 3FBE
    UIF
    GPMMPXJOH
    BOE
    EFUFSNJOF
    JG
    UIF
    IZQPUIFTJT
    DBO
    CF
    JOGFSSFE
    GSPN
    UIF
    QSFNJTF 1SFNJTF
    3VTTJBO
    DPTNPOBVU
    7BMFSZ )ZQPUIFTJT
    3VTTJBOT
    IPME
    UIF
    SFDPSE 0QUJPOT
    ZFT
    OP ճ౴Λੜ੒ ೖྗ
12. • എܠ֓ཁ • ఏҊख๏σʔληοτ • ධՁ࣮ݧ • ߟ࡯ ໨࣍ 12
13. • (15
    #
    #SPXO
    FU
    BM
     • (-B. #&
    %V
    FU
    BM
     • -B.%"15
    #
    5IPQQJMBO FU
    BM
     – XFC্ͷจষɺର࿩σʔλɺXJLJQFEJBͰ ࣄલֶशͨ͠EFDPEFSPOMZݴޠϞσϧ – ࣄલֶशίʔύεʹ͸ίʔυ΍ӳޠҎ֎ͷ จষ΋ؚ·ΕΔ • '-"/
    # – ࣄલֶशࡁΈ-B.%"15Λ JOTUSVDUJPO
    UVOJOH ࣮ݧઃఆ 13 ϕʔεϥΠϯͱఏҊ๏ σʔληοτ ϕʔε ϥΠϯ ఏҊ๏ ֶश ධՁ • ධՁͰ࢖༻͢ΔΫϥελҎ֎ͷશσʔληο τΛࠞ߹ • ֤σʔληοτͷσʔλ਺Λ࠷େສ݅ʹ੍ ݶ͠ɺσʔληοτؒͷෆۉߧΛܰݮ • ςϯϓϨʔτ͸σʔλຖʹϥϯμϜબ୒ • ԼهͷΫϥελͰධՁ – OBUVSBM
    MBOHVBHF
    JOGFSFODF
    SFBEJOH
    DPNQSFIFOTJPO
    DMPTFECPPL
    2"
    USBOTMBUJPO
    DPNNPOTFOTF
    SFBTPOJOH
    DPSFGFSFODF
    SFTPMVUJPO
    TUSVDUUPUFYU • ֤ςϯϓϨʔτͰಘͨੑೳͷฏۉΛଌఆ – ධՁࢦඪ͸BDDVSBDZ
    #-&6ͳͲ λεΫʹΑΓҟͳΔ * '-"/ͷςϯϓϨʔτΛೖྗͯ͠΋ੑೳ͕ஶ͘͠௿͍ͨΊɺ શͯͷϕʔεϥΠϯϞσϧ͸(15ͷϓϩϯϓτͰධՁ
14. • JOTUSVDUJPO
    UVOFͨ͠Ϟσϧ ͷํ͕ɺ UVOF͍ͯ͠ͳ͍Ϟσϧ ΑΓߴ͍ੑೳ • ଟ͘ͷσʔληοτͰ(15΍(-B. Λ্ճΔ – େྔͷจষΛಡΈࠐΜͰֶश͢ΔΑΓɺ JOTUSVDUJPO
    UVOJOHͷํ͕༗ޮͱࣔࠦ instruction tuning͕ػೳͨ͠λεΫ 14 (accuracy) (accuracy/F1) (accuracy) (BLEU)
15. • ҰํɺDPNNPOTFOTF
    SFBTPOJOH΍DPSFGFSFODF
    TPMVUJPOͰ͸-B.%"15ʹର͠λεΫͷΈੑೳ͕޲্ – (15΍-B.%"15ͷϓϩϯϓτ͸MBOHVBHF
    NPEFMJOHΛͦͷ··ղ͔ͤΔ΋ͷʹͳ͍ͬͯΔ ʢೖྗͱग़ྗΛͭͳ͛ΔͱࣗવͳจষʹͳΔʣҰํɺ'-"/ͷϓϩϯϓτ͸৑௕ͳͨΊͱਪଌ – ͦ΋ͦ΋ධՁʹར༻͢Δϓϩϯϓτ͸౷Ұ͢΂͖ͱࢥ͏͕ɺ'-"/ͷϓϩϯϓτΛ(15΍-BNCEB15ʹೖྗ͢Δͱ ੑೳ͕ஶ͘͠௿͍ͨΊɺ'-"/ͷϓϩϯϓτΛධՁʹ࢖༻͠ͳ͔ͬͨͱͷ͜ͱɻٯ͸ෆ໌ instruction tuning͕ػೳ͠ʹ͍͘λεΫ 15 (15 '-"/ *
    QBDLFE
    VQ
    NZ
    CFMPOHJOHT 8IBU
    JT
    UIF
    DBVTF 015*0/4  *
    XBT
    IVOUJOH
    GPS
    B
    OFX
    BQBSUNFOU  *
    XBT
    NPWJOH
    PVU
    PG
    NZ
    BQBSUNFOU ೖྗ ਖ਼ղ ग़ྗ *
    XBT
    IVOUJOH
    GPS
    B
    OFX
    BQBSUNFOU *
    XBT
    IVOUJOH
    GPS
    B
    OFX
    BQBSUNFOU *
    QBDLFE
    VQ
    NZ
    CFMPOHJOHT
    CFDBVTF ʢCOPAσʔληοτͷೖग़ྗྫɻ੨ࣈ͕ϓϩϯϓτɻʣ
16. • ӳޠͰ΋தࠃޠͰ΋ճ౴Մೳ ੒ޭͨ͠ࢦࣔͷྫ 16 • ੜ੒λεΫ΋ղ͘͜ͱ͕Ͱ͖Δ • ʮۭ͔Βόφφ͕߱ͬͯ͘Δʯ ݱ৅Λࣔ͢৽୯ޠΛ࡞੒ ʢ࣮֬ʹֶशͯ͠ͳ͍λεΫʣ
17. • จதͷO൪໨ͷ୯ޠΛ౴͑Δͱ͍͏୯७ͳλεΫ ͢Βࣦഊ͢Δྫ͕ࢄݟ ࣦഊͨ͠ࢦࣔͷྫ 17 • σϯϚʔΫޠʢ%BOJTIʣͰճ౴͞Εͨ΋ͷͷɺ ࣭໰จ͕຋༁͞Εͯ͠·ͬͨ
18. • എܠ֓ཁ • ఏҊख๏σʔληοτ • ධՁ࣮ݧ • ߟ࡯ ໨࣍ 18
19. λεΫ/σʔληοτ/ςϯϓϨʔτͷ਺ 19 • େےͰ͸ֶशʹ࢖͏λεΫΛ૿΍͢΄Ͳੑೳ޲্ – ʢՃ͑Δॱ൪΍૊Έ߹Θͤʹ΋ΑΔͷͰҰ֓ʹ͸ ݴ͑ͳ͍͕ʣҰ෦ͷλεΫ͸௥Ճͯ͠΋ੑೳ͕͞ ΄Ͳ޲্ͤͣ • σʔληοτ਺Ϋϥελ͸૿΍͢΄Ͳੑೳ͕޲্ ʢ˛ WT
    ˔ʣ • ςϯϓϨʔτ਺σʔληοτ͸ੑೳʹ͋·ΓӨڹ ͠ͳ͍ʢԣ࣠ʣ
20. • Ϟσϧ͕େ͖͘ͳΔ΄Ͳੑೳ͕޲্ɻશͯͷ࿦จͰಉ༷ͷ܏޲͕ใࠂɻ • '-"/Ͱ͸#ҎԼͷ৔߹JOTUSVDUJPO
    UVOJOHʹΑΓΉ͠Ζੑೳ͕௿Լ͢Δ͜ͱΛใࠂ͍ͯ͠Δ͕ɺ ଞͷ࿦จͰ͸ΑΓখ͍͞ϞσϧͰJOTUSVDUJPO
    UVOJOHͷ༗ޮੑΛใࠂ – 5
    /BUVSBM
    *OTUSVDUJPOT
    W
    5-.
    # – .FUB*$-
    (15
    MBSHF
    . – /BUVSBM
    *OTUSVDUJPOT
    W
    #"35
    CBTF
    . • ϓϩϯϓτͷҧ͍ʢԼͭ͸ྫࣔΛՃ͍͑ͯΔʣ ࣄલֶशͷҧ͍ʢ$BVTBM
    -.
    WT
    .BTLFE
    -.ʣ ͕ݪҼʁ Ϟσϧͷେ͖͞ 20 A100/V100 ͰֶशՄೳ
21. • ύλʔϯͷϓϩϯϓτʢۭനɺσʔληοτ໊ɺλεΫͷࢦࣔʣ ͷ͏ͪɺֶश࣌ʹʮλεΫͷࢦࣔʯΛՃ͑ͨ৔߹͕࠷΋ߴੑೳʢӈਤʣ – λεΫͷJOEJDBUPS͕͋Ε͹͍͍ͱ͍͏Θ͚Ͱ͸ͳ͍ • λεΫͷࢦࣔʹՃ͑ͯྫࣔΛೖྗͨ͠৔߹ʢ'FXTIPUʣɺ શମతʹੑೳ޲্ʢԼਤʣ – ֶशɾධՁ࣌ͷϓϩϯϓτʹݟຊΛݸϥϯμϜʹ௥Ճ – 4USVDU
    UP
    UFYUͳͲPVUQVUͷܗ͕ࣜෳࡶͳλεΫͰಛʹੑೳ޲্ʁ ϓϩϯϓτͷ಺༰ 21 ʢྫࣔΛՃ͍͑ͯͳ͍'-"/ʣ ʢྫࣔΛՃ͑ͨ'-"/ʣ
22. • /BUVSBM
    *OTUSVDUJPOTͰ͸਺छྨͷϓϩϯϓτΛ࡞੒ – %FG
    λεΫͷఆٛʢࢦࣔʣ
    1PT
    λεΫͷਖ਼ྫʢྫࣔʣ
    /FH
    λεΫͷෛྫ
    &YQM
    ਖ਼ྫͱෛྫͷࠜڌͷઆ໌ • /FHʢෛྫʣ΍&YQMʢࠜڌͷઆ໌ʣΛՃ͑ͯ΋ੑೳ͸޲্ͤͣɺ%FG 1PTʢࢦࣔʴྫࣔʣͰे෼ λεΫͷࢦࣔͱྫࣔҎ֎ʹ༗ޮͳϓϩϯϓτ͸͋Δ͔ʁ 22 ग़య: Wang et al., Benchmarking Generalization via In-Context Instructions on 1,600+ Language Tasks. arXiv, 2022. ֶश࣌ʹ ࢖༻ͨ͠ ϓϩϯϓτ ධՁ࣌ʹ࢖༻ͨ͠ϓϩϯϓτ දதͷ਺஋͸ෳ਺λεΫͷฏۉੑೳʢ306(&-ʣ
23. • ϓϩϯϓτΛՃ͑ͯϚϧνλεΫֶशΛߦ͏JOTUSVDUJPO
    UVOJOHΛఏҊ͠ɺ ϓϩϯϓτΛՃ͑ͳ͍৔߹ʹൺ΂ͯɺ൚Խੑೳ͕޲্͢Δ͜ͱΛ֬ೝ – େن໛ϞσϧͰେྔͷจষΛಡΈࠐΉํ๏ʢ(15ͳͲʣͱ͸ҟͳΔΞϓϩʔν͕ൃݟ͞Εͨ – ͲͪΒ͕༏Ε͍ͯΔ͔͸ެฏʹධՁͰ͖͍ͯͳ͍ • ൚Խੑೳʹର͠༩͑ΔӨڹ͸ҎԼͷ௨Γ – λεΫσʔληοτ਺ɿଟ͍΄͏͕ੑೳ͸ߴ·Δ͕ɺੑೳ޲্΁ͷد༩౓͸λεΫʹΑΓ·ͪ·ͪ – Ϟσϧͷେ͖͞ɿେ͖͍΄Ͳੑೳ͸ߴ͍ɻҰఆҎ্ͷେ͖͕͞ඞཁ͔͸ෆಁ໌ – ϓϩϯϓτͷ಺༰ɿݱঢ়Ͱ͸ɺλεΫͷࢦࣔͱྫࣔʢਖ਼ྫʣ͕࠷ߴੑೳ • ࠓޙͷݚڀ՝୊ – ຊ౰ʹϓϩϯϓτΛཧղͰ͖͍ͯΔ͔ʁ 
    ࣍ͷ৿Լ͞Μͷ࿦จ঺հΛָ͓͠Έʹʂ > 8FCTPO 
    1BWMJDL
    l%P
    1SPNQU#BTFE
    .PEFMT
    3FBMMZ
    6OEFSTUBOE
    UIF
    .FBOJOH
    PG
    5IFJS
    1SPNQUT z
    /""$-
     – ൚Խੑೳ޲্ʹ༗༻ͳࢦࣔ΍λεΫ͸Կ͔ʁ – ೚ҙͷࢦࣔΛཧղͰ͖Δ͜ͱΛͲ͏ධՁ͢Δ͔ʁ ݁࿦ 23
24. APPENDIX 24
25. • λεΫʢσʔληοτʣͷ਺͕૿͑Δͱੑೳͷ தԝ஋͸ߴ͘ͳΔ͕ɺ෼ࢄ΋େ͖͘ͳΓ͏Δ σʔληοτ਺/ςϯϓϨʔτ਺ɿT0Ͱ͸΍΍ҟͳΔ܏޲Λࣔ͢ 25 • ςϯϓϨʔτ਺σʔληοτΛ૿΍͢ͱ ੑೳͷதԝ஋͸্͕Δ෼ࢄ͸খ͘͞ͳΔ – ಛʹຊདྷͷλεΫͱೖग़ྗ͕ҟͳΔςϯϓϨʔτ ΛՃ͑ͨ৔߹ʹੑೳ޲্ʢ྘৭ʣ σʔληοτ਺ɿ ςϯϓϨʔτ਺ɿ