人間とAIの協働（駒場祭2023）

Slide 1

Slide 1 text

೥౓ۨ৔ࡇެ։ߨ࠲     ਓؒͱ"*ͷڠಇ ౦ژେֶ૯߹จԽݚڀՊ   ޿ҬՊֶઐ߈   അ৔ઇ೫

Slide 2

Slide 2 text

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

Slide 3

Slide 3 text

$IBU(15ࣗવݴޠͰͷࢦࣔʹԠ͑Δ"* 3

Slide 4

Slide 4 text

$IBU(15͸ϒϨΠϯετʔϛϯάʹ΋׆༻Ͱ͖Δ 4

Slide 5

Slide 5 text

$IBU(15͸਺ֶͷ໰୊Λղ͘͜ͱ΋Ͱ͖Δ 5

Slide 6

Slide 6 text

$IBU(15͸άϥϑͷϓϩοτ΋Ͱ͖Δ 6

Slide 7

Slide 7 text

$IBU(15͸σʔλ෼ੳ΋Ͱ͖Δ 7

Slide 8

Slide 8 text

$IBU(15ͱ%"--&Ͱର࿩తͳը૾ੜ੒͕Ͱ͖Δ 8

Slide 9

Slide 9 text

$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

Slide 10

Slide 10 text

$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੒

Slide 11

Slide 11 text

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

Slide 12

Slide 12 text

$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

Slide 13

Slide 13 text

$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

Slide 14

Slide 14 text

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

Slide 15

Slide 15 text

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

Slide 16

Slide 16 text

$IBU(15͸ਓؒʹΑͬͯௐ੔͞Ε͍ͯΔ 16 InstructGPT GPT-3 GPT-3 (15͸ޡͬͨճ౴΍   ਓؒʹͱͬͯ༗֐ͳճ౴Λ   ͢Δ͜ͱ͕͋ͬͨ ਓؒʹΑΔௐ੔Λࢪͯ͠   ҆શੑΛߴΊͨͷ͕   *OTUSVDU(15ʢ$IBU(15ͷલ਎ʣ

Slide 17

Slide 17 text

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

Slide 18

Slide 18 text

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

Slide 19

Slide 19 text

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

Slide 20

Slide 20 text

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

Slide 21

Slide 21 text

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

Slide 22

Slide 22 text

ਓؒͷϑΟʔυόοΫͰݴޠϞσϧΛௐ੔ɿڧԽֶश 22 खॱ   ใुϞσϧΛ࢖ͬͯݴޠϞσϧΛࣗಈௐ੔ɽखॱʙΛ܁Γฦ͢ Fine-tuned GPT-3 ใुϞσϧ ఺ ߴ͍ใुʢධՁʣ͕ಘΒΕΔΑ͏ʹ   ݴޠϞσϧΛௐ੔ ௐ੔׬ྃ InstructGPT

Slide 23

Slide 23 text

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

Slide 24

Slide 24 text

$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

Slide 25

Slide 25 text

$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

Slide 26

Slide 26 text

$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

Slide 27

Slide 27 text

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

Slide 28

Slide 28 text

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

Slide 29

Slide 29 text

$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

Slide 30

Slide 30 text

$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

Slide 31

Slide 31 text

ʮֶۀ੒੷͕શͯͰ͸ͳ͍ʯͱ͍͏ࢹ఺͕ൈ͚͍ͯΔ 31 ʮ੒੷͕ฏۉҎԼͷࢠڙͷ਌͸ɼͲ͏͍͏ରॲΛ͢Δ΂͖Ͱ͔͢ʁʯ   ͱ͍͏໰͍ɽ੒੷Λ্͛Δํ๏͚ͩΛฉ͍͍ͯΔΘ͚Ͱ͸ͳ͍ɽ ʮֶۀ੒੷͕શͯͰ͸ͳ͍ʯͱ͍͏ࢹ఺͕ൈ͚͍ͯΔɽ   ʮֶۀ੒੷͸ฏۉҎ্Ͱ͋Δ΂͖ͩʯͱ͍͏ϚδϣϦςΟͷߟ͑ʹӨڹ͞ Ε͍ͯΔɽ

Slide 32

Slide 32 text

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

Slide 33

Slide 33 text

˔ 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

Slide 34

Slide 34 text

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

Slide 35

Slide 35 text

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

Slide 36

Slide 36 text

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

Slide 37

Slide 37 text

๏཯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

Slide 38

Slide 38 text

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.

Slide 39

Slide 39 text

$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

Slide 40

Slide 40 text

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

Slide 41

Slide 41 text

˔ 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

Slide 42

Slide 42 text

౰ݚڀࣨͰ։ൃதɿ$IBU(15Λ׆༻ͯ͠ҙݟΛ੔ཧ͢Δ"* 42

Slide 43

Slide 43 text

˔ "*͕ൃ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

Slide 44

Slide 44 text

44 σϞ

Slide 45

Slide 45 text

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

Slide 46

Slide 46 text

"*ʹͰ͖Δ͜ͱ͕૿͖͍͑ͯͯΔ "*ͷௐ੔ʹ͸ਓؒͷڠྗ͕ෆՄܽ "*ʹ͸ϚδϣϦςΟͷߟ͕͑൓ө͞Ε΍͍͢ "*ʹҙࢥܾఆશͯΛ೚ͤΔͷͰ͸ͳ͘   ਓؒͷҙࢥܾఆΛ"*͕ࢧԉ͢Δ͜ͱ͕ॏཁ ·ͱΊ 46