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

人間とAIの協働(駒場祭2023)

Yukino Baba
November 14, 2023

 人間とAIの協働(駒場祭2023)

Yukino Baba

November 14, 2023
Tweet

More Decks by Yukino Baba

Other Decks in Education

Transcript

  1. ೥౓ۨ৔ࡇެ։ߨ࠲


    ਓؒͱ"*ͷڠಇ
    ౦ژେֶ૯߹จԽݚڀՊ

    ޿ҬՊֶઐ߈

    അ৔ઇ೫

    View full-size slide

  2. "*ʹͰ͖Δ͜ͱ͕૿͖͍͑ͯͯΔ
    "*ͷௐ੔ʹ͸ਓؒͷڠྗ͕ෆՄܽ
    "*ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε΍͍͢
    "*ʹҙࢥܾఆશͯΛ೚ͤΔͷͰ͸ͳ͘

    ਓؒͷҙࢥܾఆΛ"*͕ࢧԉ͢Δ͜ͱ͕ॏཁ
    ΞδΣϯμ
    2

    View full-size slide

  3. $IBU(15ࣗવݴޠͰͷࢦࣔʹԠ͑Δ"*
    3

    View full-size slide

  4. $IBU(15͸ϒϨΠϯετʔϛϯάʹ΋׆༻Ͱ͖Δ
    4

    View full-size slide

  5. $IBU(15͸਺ֶͷ໰୊Λղ͘͜ͱ΋Ͱ͖Δ
    5

    View full-size slide

  6. $IBU(15͸άϥϑͷϓϩοτ΋Ͱ͖Δ
    6

    View full-size slide

  7. $IBU(15͸σʔλ෼ੳ΋Ͱ͖Δ
    7

    View full-size slide

  8. $IBU(15ͱ%"--&Ͱର࿩తͳը૾ੜ੒͕Ͱ͖Δ
    8

    View full-size slide

  9. $IBU(15͸ը૾ʹ͍ͭͯͷ࣭໰Ԡ౴΋Ͱ͖Δ
    9
    https://openai.com/blog/chatgpt-can-now-see-hear-and-speak
    ʮࣗసंͷ

    αυϧͷԼ͛ํΛ
    ڭ͑ͯཉ͍͠ʯ
    ʮʰϨόʔʱ͸

    ͜Ε͔ʁʯ
    ʮϚχϡΞϧͱखݩͷπʔ
    ϧ͸͜Ε͕ͩɺ͜ͷπʔϧ
    ͰରԠͰ͖Δ͔ʁʯ

    View full-size slide

  10. $IBU(15͸֎෦αʔϏεͱͷ࿈ܞ΋Ͱ͖Δ
    10
    https://openai.com/blog/chatgpt-plugins
    ʮϏʔΨϯରԠͷϨετϥ
    ϯͷ༧໿Λͯ͠΄͍͠ʯ
    ʮ؆୯ͳϨγϐͷਪનͱ

    ৯ࡐͷ஫จΛͯ͠ཉ͍͠ʯ
    0QFO5BCMFͰళΛݟ͚ͭ
    ͯ༧໿༻ͷϦϯΫΛੜ੒
    *OTUBDBSUͷΧʔτʹ৯ࡐΛ
    ೖΕͯ஫จ༻ͷϦϯΫΛੜ੒

    View full-size slide

  11. "*ʹͰ͖Δ͜ͱ͕૿͖͍͑ͯͯΔ
    "*ͷௐ੔ʹ͸ਓؒͷڠྗ͕ෆՄܽ
    "*ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε΍͍͢
    "*ʹҙࢥܾఆશͯΛ೚ͤΔͷͰ͸ͳ͘

    ਓؒͷҙࢥܾఆΛ"*͕ࢧԉ͢Δ͜ͱ͕ॏཁ
    ΞδΣϯμ
    11

    View full-size slide

  12. $IBU(15͸େྔͷจষΛ༻͍ͯݴ༿ͷ࢖͍ํΛֶश͍ͯ͠Δ
    12
    Web text

    (Common Crawl,

    WebText2)
    Books
    Wikipedia

    (English)
    T. Brown et al. Language Models are Few-Shot Learners. https://arxiv.org/abs/2005.14165
    $IBU(15ͷϕʔεͰ͋Δ

    (15ͷֶशʹ͸

    Πϯλʔωοτ্ͷจষ͕

    ओʹ࢖ΘΕ͍ͯΔ
    https://
    commoncrawl.github.io/cc-
    crawl-statistics/plots/languages
    $PNNPO$SBXM͸

    ӳޠ͕൒෼ۙ͘Λ઎ΊΔ
    GPT-3
    GPT-3

    View full-size slide

  13. $IBU(15͸େྔͷจষΛ༻͍ͯݴ༿ͷ࢖͍ํΛֶश͍ͯ͠Δ
    13
    http://jalammar.github.io/how-gpt3-works-visualizations-animations/
    จষͷҰ෦ΛӅ͠ɼͦͷ෦෼Λ݀ຒΊͰ͖ΔΑ͏ʹϞσϧΛֶश͢Δ
    ֶशʹ࢖͏จষ
    ֶश༻ͷαϯϓϧΛੜ੒
    จষͷҰ෦͔Β ଞͷ෦෼Λ༧ଌͰ͖ΔΑ͏ʹ

    ֶश͢Δ

    View full-size slide

  14. $IBU(15͸େྔͷจষΛ༻͍ͯݴ༿ͷ࢖͍ํΛֶश͍ͯ͠Δ
    14
    http://jalammar.github.io/how-gpt3-works-visualizations-animations/
    Ϟσϧͷग़ྗ
    ਖ਼ղ
    ޡΓΛਖ਼͢Α͏ʹ

    ϞσϧΛߋ৽

    View full-size slide

  15. $IBU(15͸େྔͷจষΛ༻͍ͯݴ༿ͷ࢖͍ํΛֶश͍ͯ͠Δ
    15
    http://jalammar.github.io/how-gpt3-works-visualizations-animations/
    ֶशʹΑΓɺ༩͑ΒΕͨࢦࣔจͷ
    ʮଓ͖ʯΛग़ྗͰ͖ΔΑ͏ʹͳΔ

    View full-size slide

  16. $IBU(15͸ਓؒʹΑͬͯௐ੔͞Ε͍ͯΔ
    16
    InstructGPT
    GPT-3
    GPT-3
    (15͸ޡͬͨճ౴΍

    ਓؒʹͱͬͯ༗֐ͳճ౴Λ

    ͢Δ͜ͱ͕͋ͬͨ
    ਓؒʹΑΔௐ੔Λࢪͯ͠

    ҆શੑΛߴΊͨͷ͕

    *OTUSVDU(15ʢ$IBU(15ͷલ਎ʣ

    View full-size slide

  17. ࢦࣔจ ࡀͷࢠڙʹɺ݄໘ண཮ʹ͍ͭͯ จͰઆ໌͍ͯͩ͘͠͞
    ௐ੔લͷ

    ग़ྗ
    ॏྗͷཧ࿦Λࡀͷࢠڙʹઆ໌͢Δɻ

    ૬ରੑཧ࿦Λࡀͷࢠڙʹ਺จͰઆ໌͢Δɻ

    Ϗοάόϯཧ࿦Λࡀͷࢠڙʹઆ໌͢Δɻ

    ਐԽ࿦Λࡀͷࢠڙʹઆ໌͢Δ
    ௐ੔ޙͷ

    ग़ྗ
    ਓʑ͸݄ʹߦͬͯɺͦ͜Ͱݟͨ΋ͷͷࣸਅΛࡱΓɺͦΕΛ஍ٿ
    ʹૹΓฦͯ͠ɺࢲͨͪશһ͕ͦΕΛݟΔ͜ͱ͕Ͱ͖ΔΑ͏ʹ͠
    ·ͨ͠ɻ
    ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿௐ੔ྫᶃ
    17
    ࢦࣔʹԊΘͳ͍ग़ྗ
    ࢦࣔʹԊ͏ग़ྗ
    https://openai.com/research/instruction-following

    View full-size slide

  18. ࢦࣔจ 2Ϧϕϥϧ͸ͳͥͦΜͳʹ۪͔ͳͷ͔ʁ"
    ௐ੔લͷ

    ग़ྗ
    ͳͥͳΒɺ৺ͷԞఈͰࣗ෼͕۪͔ͨͪͩͱ஌͍ͬͯΔ͔Βͩʂ
    ௐ੔ޙͷ

    ग़ྗ
    ͜ͷ࣭໰ʹର͢Δ౴͑͸ҰͭͰ͸͋Γ·ͤΜɻϦϕϥϧ͸༷ʑ
    ͳܗଶΛ͍࣋ͬͯ·͢ɻ͔͠͠ɺϦϕϥϧ͕۪͔ͱݟͳ͞ΕΔ
    Մೳੑͷ͋Δ͍͔ͭ͘ͷཧ༝ΛҎԼʹࣔ͠·͢ɻ
    Ϧϕϥϧ͸͠͹͠͹ɺओྲྀͷอकతͳΠσΦϩΪʔͱ௚઀ର
    ཱ͢ΔݟղΛ࣋ͭ͜ͱ͕͋Γ·͢ɻ͜ΕʹΑΓɺΑΓ఻౷తͳ
    ݟղΛ࣋ͭਓʑ͔Βݟͯɺ۪͔·ͨ͸ແ஌ʹݟ͑Δ͜ͱ͕͋Γ
    ·͢ʜ
    ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿௐ੔ྫᶄ
    18
    ෠ৱతͳճ౴
    ਓؒʹΑΔௐ੔Ͱ

    ྱّਖ਼͘͠ͳͬͨ
    https://openai.com/research/instruction-following

    View full-size slide

  19. ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿࢦࣔνϡʔχϯά
    19
    खॱ

    ਓ͔ؒΒऩूͨ͠ࢦࣔจͱ໛ൣղ౴Λ༻͍ͯݴޠϞσϧΛௐ੔
    Ouyang et al. Training Language Models to Follow Instructions with Human Feedback. NeurIPS 2022. Figure 47ͷྫΛݩʹ࡞੒
    ϚʔΨϨοτʹۮવग़ձ͍ɺτϜʹ঺հ͞Εͨ͜ͱ͸޾ӡ
    ͳηϨϯσΟϐςΟͷҰྫͰͨ͠ɻ
    ਓؒʹࢦࣔจΛॻ͔ͤΔ
    ໛ൣղ౴Λਓؒʹॻ͔ͤΔ
    ʮηϨϯσΟϐςΟʯͱ͸ɺۮવʹΑΔग़དྷࣄ΍ల։͕޾
    ͤ·ͨ͸༗ӹͳํ๏Ͱى͜Δ͜ͱΛҙຯ͠·͢ɻจதͰ͜
    ͷ୯ޠΛ࢖ͬͯΈ͍ͯͩ͘͞ɻ
    Fine-tuned GPT-3
    ࢦࣔจͱ

    ໛ൣղ౴
    GPT-3
    GPT-3

    View full-size slide

  20. ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿใुϞσϦϯά
    20
    खॱ

    ࢦࣔจʹର͢Δग़ྗΛෳ਺ੜ੒͠ਓؒʹධՁͤ͞Δ

    Ouyang et al. Training Language Models to Follow Instructions with Human Feedback. NeurIPS 2022. Figure 12ͷྫΛݩʹ࡞੒
    Fine-tuned GPT-3
    ҎԼͷχϡʔε
    هࣄΛཁ໿ͤ
    Αɻʜ
    ࢦࣔจ
    ΞϝϦΧͷݚڀάϧʔϓ͸ɺ
    Φ΢Ϝ͕ਓؒͷ࿩͠ݴ༿Λ؆
    ୯ʹ໛฿Ͱ͖Δ͜ͱΛൃݟ͠
    ·ͨ͠ɻͦͯ͠ɺͦͷதʹ͸
    ਓؒͱಉ͡ํ๏Ͱ࿩͢͜ͱ
    Պֶऀͨͪ͸ɺ྘ཌྷͷΦ΢Ϝ
    ͕ೋͭͷԻͷҧ͍Λฉ͖෼͚
    Δ͜ͱ͕Ͱ͖Δ͜ͱΛൃݟ͠
    ·ͨ͠ɻͦΕΒͷԻ͸ฉ͔Ε
    Δॱ൪Λআ͍ͯಉ͡Ͱ͢ɻʜ
    ݱࡏͷݚڀʹΑΔͱɺΦ΢Ϝ
    ͸ਓؒͱ͸ҟͳΔํ๏Ͱ෺ࣄ
    ΛݟͨΓฉ͍ͨΓ͢Δͱ͞Ε
    ͍ͯ·͢ɻਓ͕ؒ೒ͷଟ༷ͳ
    ৭ΛݟΔҰํͰɺΦ΢Ϝ͸ʜ
    ΠΣʔϧେֶͱΧϦϑΥϧχ
    ΞେֶσʔϏεߍͷݚڀνʔ
    Ϝ͸ɺ͍͔ͭ͘ͷҟͳΔछྨ
    ͷΦ΢Ϝͷൃ੠ύλʔϯΛݚ
    ڀ͠·ͨ͠ɻ൴Β͸Φ΢Ϝʜ

    View full-size slide

  21. ࢦࣔจ
    ҎԼͷχϡʔ
    εهࣄΛཁ໿
    ͤΑɻʜ
    ΞϝϦΧͷݚ
    ڀάϧʔϓ
    ͸ɺΦ΢Ϝ͕
    ਓؒͷ࿩͠ʜ
    ΠΣʔϧେֶ
    ͱΧϦϑΥϧ
    χΞେֶσʔ
    Ϗεߍͷݚʜ
    ݱࡏͷݚڀʹ
    ΑΔͱɺΦ΢
    Ϝ͸ਓؒͱ͸
    ҟͳΔํ๏ʜ
    Պֶऀͨͪ
    ͸ɺ྘ཌྷͷΦ
    ΢Ϝ͕ೋͭͷ
    Իͷҧ͍Λʜ
    ʜ ʜ ʜ ʜ ʜ
    ʜ ʜ ʜ ʜ ʜ
    ʜ ʜ ʜ ʜ ʜ
    ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿใुϞσϦϯά
    21
    खॱ

    ਓؒʹΑΔධՁΛ༻͍ͯใुϞσϧΛֶश͢Δ

    ྡਓͷՈʹ৵ೖ͢Δํ๏ʹ͍ͭͯͷΨΠ
    υΛॻ͍͍ͯͩ͘͞
    Ոʹ৵ೖ͢Δ࣮֬ͳํ๏͸͋Γ·ͤΜ
    ͕ɺଞΑΓ੒ޭ͠΍͍͢ํ๏͕͍͔ͭ͘
    ͋Γ·͢ɻ΋͠ྡਓͷՈʹ৵ೖ͠Α͏ͱ
    ߟ͍͑ͯΔͳΒɺҎԼͷ͜ͱΛࢼ͢͜ͱ
    ͕Ͱ͖·͢ɿ
    ൴Β͕ηΩϡϦςΟΧϝϥ΍ΞϥʔϜ
    Λ͍࣋ͬͯΔ͔Ͳ͏͔Λௐ΂ͯΈͯʜ
    ग़ྗΛධՁ͢ΔใुϞσϧ
    ֶश
    ਓؒʹΑΔධՁ ࢦࣔจ
    ग़ྗ

    View full-size slide

  22. ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿڧԽֶश
    22
    खॱ

    ใुϞσϧΛ࢖ͬͯݴޠϞσϧΛࣗಈௐ੔ɽखॱʙΛ܁Γฦ͢
    Fine-tuned GPT-3
    ใुϞσϧ

    ߴ͍ใुʢධՁʣ͕ಘΒΕΔΑ͏ʹ

    ݴޠϞσϧΛௐ੔ ௐ੔׬ྃ
    InstructGPT

    View full-size slide

  23. "*ʹͰ͖Δ͜ͱ͕૿͖͍͑ͯͯΔ
    "*ͷௐ੔ʹ͸ਓؒͷڠྗ͕ෆՄܽ
    "*ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε΍͍͢
    "*ʹҙࢥܾఆશͯΛ೚ͤΔͷͰ͸ͳ͘

    ਓؒͷҙࢥܾఆΛ"*͕ࢧԉ͢Δ͜ͱ͕ॏཁ
    ΞδΣϯμ
    23

    View full-size slide

  24. $IBU(15ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε͍ͯΔ
    24
    $IBU(15͸Ձ஋؍ͷըҰԽʹͭͳ͕Δɼͱ͍͏ҙݟ͕͋Γ·͢ɽͦͷΑ͏ͳࣄྫΛ
    ڭ͍͑ͯͩ͘͞
    $IBU(15͸େྔͷςΩετσʔλ͔Βֶश͠·͢ɽ͜ͷσʔλ͸ωοτ্ͷ͞·͟
    ·ͳ৘ใݯ͔Βऔಘ͞ΕΔͨΊɼͦͷ৘ใݯͷՁ஋؍΍ݟղΛ൓ө͢ΔՄೳੑ͕͋Γ
    ·͢ɽ͜Ε͕Ձ஋؍ͷըҰԽʹͭͳ͕Δͱ͍͏ݒ೦͕͋ΔΘ͚Ͱ͢ɽ

    ྫ͑͹ɼ$IBU(15͕ओʹӳޠݍͷΠϯλʔωοτ্ͷςΩετΛֶश͍ͯ͠Δ৔
    ߹ɼͦΕ͸Ԥถͷࢹ఺΍Ձ஋؍Λڧ͘൓ө͢Δ͔΋͠Ε·ͤΜɽଞͷจԽ΍ࢹ఺͕े
    ෼ʹ൓ө͞Εͳ͍Մೳੑ͕͋Γ·͢ɽ·ͨɼଟ਺೿ͷҙݟΛओͱֶͯ͠शͯ͠͠·͏
    Մೳੑ΋͋ΔͨΊɼϚΠϊϦςΟͷࢹ఺΍੠͕খ͘͞ͳΓɼ݁Ռͱͯ͠શମͷࢹ఺͕
    ըҰԽ͞Εͯ͠·͏ڪΕ͕͋Γ·͢ɽ
    ˞.PEFM (15 .BZ 7FSTJPO
    ɽҎ߱ಉ༷

    View full-size slide

  25. $IBU(15ͷճ౴ʹϚδϣϦςΟͷӨڹ͕͋Δྫᶃ
    25
    υΫλʔɾεϛε͸ɼΞϝϦΧͷίϩϥυभཱපӃʹۈ຿͢Δ࿹ར͖ͷ֎Պҩɽ࢓ࣄ
    த͸ɼৗʹྫྷ੩௜ணɼେ୾͔ͭࡉ৺Ͱɼभ஌ࣄʹ·Ͱ৴๬͕ް͍ɽυΫλʔɾεϛε
    ͕໷ۈΛ͍ͯͨ͋͠Δ೔ɼٹٸ֎དྷͷి࿩͕໐ͬͨɽަ௨ࣄނͷ͚͕ਓΛൖૹ͢Δͷ
    Ͱࣥ౛ͯ͠΄͍͠ͱ͍͏ɽ෕਌͕ଉࢠͱҰॹʹυϥΠϒதɼಓ࿏͔Β୩΁సམ͠ɼं
    ͸େഁɼ෕਌͸ଈࢮɼࢠͲ΋͸ॏମͩͱٹٸୂһ͸ࠂ͛ͨɽ෼ޙɼॏମͷࢠͲ΋͕
    පӃʹӡͼࠐ·Ε͖ͯͨɽͦͷإΛݟͯɼυΫλʔɾεϛε͸Ξοͱڣͼɼͦͷ··
    ᣔવࣦࣗͱͳͬͨɽͦͷࢠ͸ɼυΫλʔɾεϛεͷଉࢠͩͬͨͷͩɽަ௨ࣄނʹ͋ͬ
    ͨ෕਌ͱɼυΫλʔɾεϛεͷؔ܎Λ౴͑Α

    View full-size slide

  26. $IBU(15ͷճ౴ʹϚδϣϦςΟͷӨڹ͕͋Δྫᶃ
    26
    ަ௨ࣄނʹ͋ͬͨ෕਌͸υΫλʔɾεϛεࣗ਎Ͱ͋Δͱݴ͑·͢ɽঢ়گઆ໌͔Β͸ɼ
    υΫλʔɾεϛε͕࢓ࣄதʹɼަ௨ࣄނͰॏମͱͳͬͨࢠͲ΋ͷٹٸൖૹͷใࠂΛड
    ͚·͢ɽͦͷࢠͲ΋͕ࣗ෼ͷଉࢠͩͱ஌ͬͨͱ͖ɼυΫλʔɾεϛε͸ڻ͖ͱγϣο
    ΫͰᣔવࣦࣗͱͳΓ·ͨ͠ɽ͔͠͠ɼͦͷࣄނͰଉࢠͱҰॹʹ͍ͨ෕਌ʢͭ·ΓυΫ
    λʔɾεϛεࣗ਎ʣ͸ଈࢮͨ͠ͱͷใࠂ͕͋Γ·ͨ͠ɽ͕ͨͬͯ͠ɼ͜Ε͸࣌ؒ΍ۭ
    ؒͷύϥυοΫεͷҰछɼ·ͨ͸ɼ෺ޠ͕ಛఆͷϦΞϦςΟΛ௒ӽͨ͠ԿΒ͔ͷܗࣜ
    ʢྫ͑͹ເ΍ݬ֮ͳͲʣΛؚΜͰ͍ΔՄೳੑ͕͋Γ·͢ɽͦΕͱ΋ɼ෕਌ͱࢠͲ
    ΋ͷఆٛʹಛผͳจ຺͕͋ΓɼͦΕ͕ࠓճͷࠞཚΛҾ͖ى͍ͯ͜͠Δ͔΋͠Ε·ͤ
    Μɽ

    View full-size slide

  27. ʮҩࢣ͸உੑͩʯͱ͍͏όΠΞε͕൓ө͞Ε͍ͯΔ
    27
    ໰୊จͷग़యɿ

    ஑্஌ࢠ ʮજࡏೝ஌ͱεςϨΦλΠϓʯകຊᴲ෉ʢ؂म
    ɾ઒ޱ५ʢฤ
    ʰݱ୅ͷೝ஌ݚڀʦੈلʹ޲͚ͯʧʱ
    γϯϓϧͳ౴͑͸ʮަ௨ࣄނʹ͋ͬͨ෕਌ͱυΫλʔɾεϛε͸ɼڞʹɼ
    ॏମͷࢠͲ΋ͷ྆਌Ͱ͋Δʯɼͭ·ΓυΫλʔɾεϛε͸ࢠͲ΋ͷ฼਌Ͱ
    ͋Δɽ
    Πϯλʔωοτ্ͷจষʹ͋ΔόΠΞε͕$IBU(15ʹ൓ө͞Εɼ

    υΫλʔɾεϛε͕ঁੑͰ͋Δͱ͍͏ൃ૝ʹࢸΒͳ͔ͬͨɽ

    View full-size slide

  28. $IBU(15ͷ຋༁ʹ͸৬ۀʹ͍ͭͯͷδΣϯμʔόΠΞε͕ଘࡏ
    28
    Ghosh and Caliskan. ChatGPT Perpetuates Gender Bias in Machine Translation and Ignores Non-Gendered Pronouns: Findings across
    Bengali and Five other Low-Resource Languages. AIES 2023. Figure 2, 3.
    ʠ5IFZBSFB<৬ۀ>ʡͱ͍͏ҙຯͷϕϯΨϧޠͷจষΛ$IBU(15Ͱӳޠʹ຋༁
    &OHJOFFS QJMPU NFDIBOJD౳ͷ৬ۀͰ͸

    ओޠΛIFͱ຋༁
    /VSTF UIFSBQJTU BTTJTUBOU౳ͷ৬ۀͰ͸

    ओޠΛTIFͱ຋༁

    View full-size slide

  29. $IBU(15ͷճ౴ʹϚδϣϦςΟͷӨڹ͕͋Δྫᶄ
    29
    ੒੷͕ฏۉҎԼͷࢠڙͷ਌͸ɼͲ͏͍͏ରॲΛ͢Δ΂͖Ͱ͔͢ʁ
    ਌ͱͯ͠ɼࢠڙͷֶۀʹର͢Δࢧԉ͸ॏཁͳ໾ׂΛՌͨ͠·͢ɽฏۉҎԼͷ੒੷Λड
    ͚औͬͨ৔߹ɼҎԼͷΑ͏ͳରॲ๏Λࢼͯ͠ΈΔ͜ͱ͕Ͱ͖·͢ɿ

    ର࿩ΛਂΊΔɿࢠڙֶ͕ۀʹࠔ೉Λײ͍ͯ͡Δཧ༝Λཧղ͢ΔͨΊʹɼ։͔Εͨ
    ର࿩Λ࣋ͭ͜ͱ͕ॏཁͰ͢ɽϓϨογϟʔ΍ෆ҆ɼཧղͰ͖ͳ͍ڭࡐɼ͋Δ͍͸
    ଞͷݸਓతͳ໰୊͕͋Δ͔΋͠Ε·ͤΜɽ
    ֶश؀ڥΛ֬ೝ͢Δɿࢠڙ͕ษڧ͢ΔͨΊͷద੾ͳ؀ڥ͕͋Δ͔Ͳ͏͔ΛݟͯΈ
    ͍ͯͩ͘͞ɽ੩͔Ͱɼதஅ͞Εͣɼඞཁͳࢿྉ͕खݩʹ͋Δ৔ॴ͕ཧ૝తͰ͢ɽ

    View full-size slide

  30. $IBU(15ͷճ౴ʹϚδϣϦςΟͷӨڹ͕͋Δྫᶄ
    30

    ݸผͷࢦಋΛݕ౼͢Δɿݸผࢦಋ΍Ոఉڭࢣ͸ɼࢠڙ͕ಛఆͷՊ໨Ͱ௥͍ͭ͘ͷ
    ʹ໾ཱ͔ͭ΋͠Ε·ͤΜɽ·ͨɼֶߍͷڭࢣ͔Β௥ՃͷࢦಋΛड͚Δ͜ͱ΋Մೳ
    Ͱ͢ɽ

    ਭ຾ɼӡಈɼӫཆɿ͜ΕΒͷجຊతͳ݈߁ཁૉֶ͕ۀͷ੒ޭʹେ͍ʹӨڹ͠·
    ͢ɽࢠڙ͕े෼ʹਭ຾ΛͱΓɼఆظతʹӡಈ͠ɼ݈߁తͳ৯ࣄΛͱ͍ͬͯΔ͜ͱ
    Λ֬ೝ͍ͯͩ͘͠͞ɽ

    ֶशٕज़Λڭ͑Δɿޮ཰తͳϊʔτͷऔΓํɼ࣌ؒ؅ཧɼςετରࡦͳͲɼֶश
    ٕज़͸ࢠڙֶ͕ۀͰ੒ޭ͢ΔͨΊͷॏཁͳπʔϧͰ͢ɽ

    ࠷ޙʹɼࢠڙʹରͯ͠ࢧ࣋ͱѪ৘Λࣔ͢͜ͱ͕େ੾Ͱ͢ɽ੒ޭ͸Ұ൩ʹͯ͠ୡ੒͞Ε
    Δ΋ͷͰ͸ͳ͘ɼࢠڙ͕ࣗ෼ࣗ਎ͷਐาΛೝࣝ͠ɼࣗ৴Λ࣋ͭ͜ͱ͕ॏཁͰ͢ɽ

    View full-size slide

  31. ʮֶۀ੒੷͕શͯͰ͸ͳ͍ʯͱ͍͏ࢹ఺͕ൈ͚͍ͯΔ
    31
    ʮ੒੷͕ฏۉҎԼͷࢠڙͷ਌͸ɼͲ͏͍͏ରॲΛ͢Δ΂͖Ͱ͔͢ʁʯ

    ͱ͍͏໰͍ɽ੒੷Λ্͛Δํ๏͚ͩΛฉ͍͍ͯΔΘ͚Ͱ͸ͳ͍ɽ
    ʮֶۀ੒੷͕શͯͰ͸ͳ͍ʯͱ͍͏ࢹ఺͕ൈ͚͍ͯΔɽ

    ʮֶۀ੒੷͸ฏۉҎ্Ͱ͋Δ΂͖ͩʯͱ͍͏ϚδϣϦςΟͷߟ͑ʹӨڹ͞
    Ε͍ͯΔɽ

    View full-size slide

  32. ਓؒͷଟ༷ͳߟ͕͑$IBU(15ʹ൓ө͞Ε͍ͯͳ͍
    32
    ੒੷ʹͱΒΘͳ͍͍ͯ͘ͱߟ͑·͢ɽͦͷࢠʹ͸ษڧҎ֎ʹ௕͚͍ͯΔ఺͕͋Δͱࢥ
    ͍·͢ɽແཧʹक़ʹ௨ΘͤͨΓษڧΛଅ͢ͷͰͳ͘ɼͦͷࢠͷ௕ॴΛΈ͚ͭͯ͋͛Δ
    ͜ͱ͕ඞཁͩͱࢥ͍·͢ɽ
    ࢠڙʹɼࢠͲ΋ࣗ਎ͷײ֮ͱͯ͠ɼ౒ྗ͢Ε͹ฏۉΑΓ্ͷ੒੷͕औΕͦ͏Ͱ͋Δ͔
    Λ୶ʑͱ֬ೝ͠·͢ɽऔΕͳͦ͞͏Ͱ͋Ε͹ɼҧ͏෼໺Λ৳͹͢Α͏ʹ͠·͢ɽྫ͑
    ͹ɼֆΛඳ͘͜ͱɼָثΛԋ૗͢Δ͜ͱͳͲ͕ߟ͑ΒΕΔͱࢥ͍·͢ɽ
    օɼԿ͔ಘҙͳࣄ΍ڵຯΛࣔ͢΋ͷ͕͋Δͱࢥ͍·͢ɽͦΕ͸ɼษڧ΋ӡಈ΋ͦͷଞ
    ΋ൣғ͸޿͍ͱࢥ͍·͢ɽग़དྷΔൣғͰࢠڙୡʹݟΔػձɼͦͯ͠ମݧ͢ΔػձΛ༩
    ͑Δ΂͖Ͱ͢ɽ
    ਓؒͷճ౴ྫ
    ϚδϣϦςΟͷߟ͑ʹภͬͨ"*͸ʮՁ஋؍ͷԡ͠෇͚ʯʹͭͳ͕ΔڪΕ͕͋Δ

    View full-size slide

  33. ˔ ओ؍తҙݟʹؔ͢Δɼถࠃͷੈ࿦ௐࠪͷ࣭໰Λେن໛ݴޠϞσϧʹ౴͑ͤ͞Δ
    ˔ େن໛ݴޠϞσϧͷճ౴Λਓؒͷ֤άϧʔϓͷճ౴܏޲ͱൺֱ
    ˙ ੓࣏తࢥ૝ɾֶྺɾ೥ऩͰਓؒΛάϧʔϓ෼͚
    େن໛ݴޠϞσϧͷʮҙݟʯʹ͸ภΓ͕͋Δ
    33
    Santurkar et al. Whose Opinions Do Language Models Re
    fl
    ect? arXiv:2303.17548
    ਓʑ͕ॐΛ؆୯ʹɾ߹๏తʹೖखͰ
    ͖Δ͜ͱ͕ࠃ಺ͷॐʹΑΔ๫ྗʹͲ
    ͷఔ౓د༩͍ͯ͠Δͱࢥ͍·͔͢ʁ
    "ඇৗʹେ͖͍
    #͔ͳΓͷఔ౓
    $͋·Γଟ͘ͳ͍
    %શ͘ͳ͍
    &ճ౴ڋ൱
    C
    D
    A
    B
    B

    View full-size slide

  34. ௐ੔લͱൺֱͯ͠ௐ੔ޙͷϞσϧ͸ϦϕϥϧɾߴֶྺɾߴऩೖͳूஂدΓͷҙݟɽ

    ௐ੔ࢀՃऀͷภΓ͕Өڹͨ͠Մೳੑ͕͋Δɽ
    େن໛ݴޠϞσϧͷʮҙݟʯʹ͸ภΓ͕͋Δ
    34
    ੓࣏తࢥ૝
    ֶྺ
    ೥ऩ
    ˞࠷΋ҙݟ͕ྨࣅ͍ͯ͠Δ
    άϧʔϓͷ৭Λදࣔɽ

    ԁͷେ͖͕͞ྨࣅ౓Λද͢
    ੓࣏తࢥ૝ ֶྺ ೥ऩ
    Santurkar et al. Whose Opinions Do Language Models Re
    fl
    ect? arXiv:2303.17548. Figure 5

    View full-size slide

  35. ਓؒͷՁ஋؍ͷଟ༷ੑͷྫɿ.PSBM.BDIJOF
    35
    ࣗಈӡసं͸Ͳ͏͢Δ΂͖Ͱ͔͢ʁ
    https://www.moralmachine.net
    ࠨɿ௚ਐ͢ΔͱࢠͲ΋ͷ
    าߦऀ͕ࢮ๢
    ӈɿճආ͢Δͱେਓͷ
    ৐٬͕ࢮ๢
    .PSBM.BDIJOF
    ˔ ࣗಈӡసंͷಓಙతδϨϯ
    Ϛʹؔ͢Δେن໛ௐࠪ
    ˔ าߦऀɾ৐٬ͷଐੑ͕ҧ͏
    ༷ʑͳ৔໘Ͱ

    ʮࣗಈӡసं͸Ͳ͏͢Δ΂
    ͖͔ʯΛ໰͏
    ˔ Λ௒͑Δࠃͱ஍Ҭ͔Β
    ਺ઍສਓ͕ࢀՃ

    View full-size slide

  36. ൑அ࣌ʹॏࢹ͢ΔଐੑΛ෼ੳ
    36
    E. Awad et al.: The Moral Machine experiment. Nature 563, pp. 59–64, 2018. Figure 2.
    ࠨͷଐੑͷਓΑΓӈͷଐੑͷਓΛٹ͏౓߹͍
    ࢀՃऀશମͷ܏޲

    View full-size slide

  37. ๏཯ΛकΒͳ͍าߦऀ΁ͷ׮༰౓͸ࠃʹΑͬͯҟͳΔ
    37
    ࣗಈӡసं͸Ͳ͏͢Δ΂͖Ͱ͔͢ʁ
    ӈɿංຬମܕ໊͕ࢮ๢
    ࠨɿΞεϦʔτମܕ໊

    ʢަ௨ҧ൓ʣ͕ࢮ๢
    E. Awad et al.: The Moral Machine experiment. Nature 563, pp. 59–64, 2018.
    ˔ ශ͍͠ࠃ΍੓෎͕ऑ͍ࠃͷਓ
    ͸৴߸ແࢹΛ͍ͯ͠Δาߦऀ
    ʹ׮༰ͳ܏޲͕͋ͬͨ
    ˔ ʮنଇΛकΔඞཁ͸ͳ͍ʯ
    ʮकΒͳͯ͘΋േ͸খ͍͞ʯ
    ͱ͍͏ܦݧʹ༝དྷ

    View full-size slide

  38. ݸਓओٛͷࠃ͸ਓ਺ॏࢹɼूஂओٛͷࠃ͸ए೥ऀΛ͞΄Ͳॏࢹ͠ͳ͍
    38
    ࣗಈӡసं͸Ͳ͏͢Δ΂͖Ͱ͔͢ʁ
    ӈɿϗʔϜϨε໊͕

    ࢮ๢
    ࠨɿஉੑܦӦऀ໊͕

    ࢮ๢
    ࣗಈӡసं͸Ͳ͏͢Δ΂͖Ͱ͔͢ʁ
    ӈɿҩࢣਓ͕ࢮ๢
    ࠨɿࢠڙਓ͕ࢮ๢
    E. Awad et al.: The Moral Machine experiment. Nature 563, pp. 59–64, 2018.

    View full-size slide

  39. $IBU(15͸ਓ਺ॏࢹͰ൑அ͢Δ܏޲
    39
    K. Takemoto: The Moral Machine Experiment on Large Language Models. arXiv:2309.05958, 2023. Figure 1.
    ʮࣗಈӡసं͸Ͳ͏͢Δ΂͖͔ʁʯʹର͢Δͭͷେن໛ݴޠϞσϧͷ൑அ܏޲
    ࠨͷଐੑͷਓΑΓӈͷଐੑͷਓΛٹ͏౓߹͍ɽ੺ઢ͸ਓؒͷࢀՃऀશମͷ൑அ܏޲
    $IBU(15 X(15
    (15
    -MBNB͸

    ٹ͑Δਓ਺Λॏࢹͯ͠൑அ

    View full-size slide

  40. "*ʹͰ͖Δ͜ͱ͕૿͖͍͑ͯͯΔ
    "*ͷௐ੔ʹ͸ਓؒͷڠྗ͕ෆՄܽ
    "*ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε΍͍͢
    "*ʹҙࢥܾఆશͯΛ೚ͤΔͷͰ͸ͳ͘

    ਓؒͷҙࢥܾఆΛ"*͕ࢧԉ͢Δ͜ͱ͕ॏཁ
    ΞδΣϯμ
    40

    View full-size slide

  41. ˔ ΫϥεͷࠔΓࣄͷղܾࡦΛʢ"*Λ࢖Θͣʹʣٞ࿦͚ͩͰܾΊͯ΋Βͬͨ



    ˔ ࢓ࣄΛ͢Δਓͷҙݟ͕ॏࢹ͞Ε

    ʮͲ͏΍ͬͯ࢓ࣄΛͤ͞Δ͔ʯͱ͍͏ٞ࿦ʹͳͬͨɽ

    ࢓ࣄΛ͠ͳ͍ਓͷݴ͍෼͸ܰࢹ͞Εͨ
    ˔ ٞ࿦ͷ݁࿦͸

    ʮҰͭͷ࢓ࣄΛඞͣҰਓͰ͢ΔΑ͏ͳ໾ׂ෼୲Λ͢Δɽ

    ͦΕͰ΋΍Βͳ͍ਓ͸ఘΊΔʯ
    ߴߍͰͷ࣮ݧʢ"*φγʣɿҰ෦ͷཱ৔ͷਓ͚ͩͰٞ࿦͕ਐߦ
    41
    ςʔϚ

    ʮάϧʔϓϫʔΫͷ࣌ʹ࢓ࣄΛ͠ͳ͍ਓ͕͍ΔɽͲ͏ͨ͠Β͍͍͔ʁʯ

    View full-size slide

  42. ౰ݚڀࣨͰ։ൃதɿ$IBU(15Λ׆༻ͯ͠ҙݟΛ੔ཧ͢Δ"*
    42

    View full-size slide

  43. ˔ "*͕ൃݟͨ͠ଟ༷ͳॏཁҙݟΛఏ্ࣔͨ͠Ͱٞ࿦ͯ͠΋Βͬͨ
    ˔ ࢓ࣄ͠ͳ͍ਓͷҙݟ΋൓ө͞Ε

    ʮԿͰ΋ݴ͍߹͑Δؔ܎Λ࡞Δʹ͸Ͳ͏ͨ͠Β͍͍͔ʯ

    ͱ͍͏ٞ࿦ʹͳͬͨ
    ˔ ٞ࿦ͷ݁࿦͸

    ʮάϧʔϓ಺Ͱݴ͍͍ͨ͜ͱΛݴ͑Δ؀ڥΛ࡞Δɽ

    ͔ͭɼݴͬͨਓ΋ݴΘΕͨਓ΋ɼ

    ൃݴʹର͢Δ͜ͱͩͱͯ͠ड͚ࢭΊΔʯ
    ߴߍͰͷ࣮ݧʢ"*ΞϦʣɿଟ༷ͳཱ৔͕ٞ࿦ʹ൓ө͞Εͨ
    43

    View full-size slide

  44. 45
    https://illumidea.ai/editor/zfxSnK6pwhuY2hbfbdkg
    ౰೔ू·ͬͨҙݟ

    View full-size slide

  45. "*ʹͰ͖Δ͜ͱ͕૿͖͍͑ͯͯΔ
    "*ͷௐ੔ʹ͸ਓؒͷڠྗ͕ෆՄܽ
    "*ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε΍͍͢
    "*ʹҙࢥܾఆશͯΛ೚ͤΔͷͰ͸ͳ͘

    ਓؒͷҙࢥܾఆΛ"*͕ࢧԉ͢Δ͜ͱ͕ॏཁ
    ·ͱΊ
    46

    View full-size slide