機械学習でヒトの行動を変える

ػցֶशͰώτͷߦಈΛม͑Δ !౦ژେֶ޻ֶ෦߸ؗ ໼૔େເʢ΍͘ΒͻΖΉʣ

w ໼૔େເʢ΍͘ΒͻΖΉʣ w υΠπɾϚοΫεϓϥϯΫਓؒ։ൃݚڀॴͰ )VNBO"*ΠϯλϥΫγϣϯͷݚڀʹैࣄ ࣗݾ঺հ

w ໼૔େເʢ΍͘ΒͻΖΉʣ w (PPHMFٴͼ.JDSPTPGU3FTFBSDI1I%'FMMPX w +45"$59 ະདྷࣾձ૑଄ࣄۀ౳ͰσβΠϯࢧԉ͔Β ൃୡো͕͍ࣇࢧԉ·ͰػցֶशͷԠ༻Λ֦͛Δݚڀʹैࣄ w
ઌ݄ΑΓ+45͖͕͚͞ʮࣾձมֵج൫ʯʹͯ ʮػցֶश࣌୅ͷࣾձม༰Λཧղ͢Δ ج൫Ξϓϩʔνͷ૑ग़ʯͱ͍͏՝୊Λ։࢝ ࣗݾ঺հ

ݚڀϓϩδΣΫτྫ w Իָܦݧͷͳ͍ਓͰ΋ࣗ༝ͳදݱΛੜΈग़ͤΔॳ৺ऀ޲͚࡞ۂࢧԉ "* ͍ΖΜͳόϦΤʔγϣϯΛ ୳ࡧ͠ͳ͕Β޷͖ͳԻָʹ ग़ձ͑ΔΠϯλϥΫγϣϯ Yakura, et
al. ISMIR 2023.

ݚڀϓϩδΣΫτྫ w $07*% ԼͷΞΠυϧͷѲखձͷมԽ͔Β ςϨίϛϡχέʔγϣϯͷड༰ͱݶքΛٞ࿦ Yakura,. ACM CHI 2021.
workplace communication [1,31] ef f iciency connectedness intimate relationships [24,50] Idol groups ? The image is taken from a YouTube video only for presentation purpose. https://www.youtube.com/watch?v=6yOxmVoAE14 The photos are taken from Wikimedia Commons and Flickr under the CC-BY-SA 2.0 license. https://w.wiki/3AF4 https://www. f lickr.com/photos/kanesue/47545700921/

ݚڀͷڵຯ ػցֶशͷ׆༻ͷͨΊͷ ΠϯλϥΫγϣϯݚڀ ػցֶशٕज़ͷݚڀ Ϣʔβ ϦεΫ΍ ݶք ࣾձͰͷ ड༰
w ΋ͱ΋ͱ͸"EWFSTBSJBM&YBNQMFͳͲػցֶशηΩϡϦςΟͷྖҬΛݚڀ w ಺ࡏతͳݶքΛલఏʹɺࣾձͰػցֶशͷԠ༻৔໘Λ޿͛Δʹ͸ʁͱ͍͏ ؍఺͔ΒػցֶशͷϥετϫϯϚΠϧ໰୊Λղ͘ΠϯλϥΫγϣϯΛݕ౼ """* *+$"*

.JOEMFTT"UUSBDUPSػցֶशͷِཅੑΛ ౿·͑ͨௌ֮ϑΟʔυόοΫσβΠϯ 7 Riku Arakawa*1, Hiromu Yakura*2,3 *: equal contribution
1 : Carnegie Mellon University 2: Max-Planck Institute for Human Development (Previously: University of Tsukuba) 3: National Institute of Advanced Industrial Science and Technology (AIST)

ػցֶशٕज़ͷ׆༻Ͱͦ͏ͨ͠৔໘ΛݮΒ͢͜ͱ͸Ͱ͖ͳ͍͔ʁ S. Whittaker Rethinking Video as a Technology for Interpersonal
Communications: Theory and Design Implications. Int. J. Hum. Comput. 1995. C. Coombs. Will COVID-19 Be the Tipping Point for the Intelligent Automation of Work? A Review of the Debate and Implications for Research. Int. J. Inf. Manage. 2020. A. Kuzminykh, et al. Classification of Functional Attention in Video Meetings. ACM CHI 2020. R. S. Oeppen, et al. Human Factors Recognition at Virtual Meetings and Video Conferencing: How to Get the Best Performance From Yourself and Others. Br. J. Oral Maxillofac. Surg. 2020. w ಛʹ$07*%Ҏ߱ɺΦϯϥΠϯ΍ ಈըϕʔεͷίϛϡχέʔγϣϯͷ ػձ͸֨ஈʹ૿͍͑ͯΔ<8IJUUBLFS> w ͔͠͠ɺͦ͏͍ͬͨ৔໘Ͱਓؒ͸ ༰қʹूதΛࣦ͍͕ͪ<,J[NJOZLI > w εϚϗΛ৮ͬͨΓɺ8FCϒϥ΢δϯάΛͨ͠Γͱ͍͏ӨڹͰ ΦϯϥΠϯձٞͷੜ࢈ੑ͕௿͘ͳ͍ͬͯΔ<0QQFO > 8

w إը૾͔Β஫ҙ͕ࢄອʹͳͬͨॠؒΛݕग़͢ΔػցֶशϞσϧ͸ଘࡏ w ΦϯϥΠϯतۀͰूத͍ͯ͠ͳֶ͍ੜʹΞϥʔτ͢Δ͜ͱ͸Մೳ<5IPNBT > w ͔͠͠ɺޡݕ஌ͰΞϥʔτͯ͠͠·͏ͱٯʹूதΛଛͳ͏͚ͩͰͳ͘ γεςϜ΁ͷෆ৴ײΛট͘͜ͱʹͳΔ<%JFUWPSTU >
w ͔͠΋ɺ ҙཉͷ௿͍ͱ͖ʹʮूத͍ͯͩ͘͠͞ʯͳͲͱ͚ͩ Ξϥʔτͯ͠΋ɺͦΕͰूத͕໭Δͱ͸ݶΒͳ͍ w ਓؒͷೝ஌ϓϩηεʹґڌͨ͠ϑΟʔυόοΫσβΠϯΛ͠ͳ͍ͱ ػցֶशٕज़͚ͩͰߦಈม༰Λى͜͢͜ͱ͸೉͍͠ A. Gupta, et al. DAiSEE: Towards User Engagement Recognition in the Wild. arXiv, 2016. C. Thomas, et al. Predicting Student Engagement in Classrooms Using Facial Behavioral Cues ACM CHI MIE Workshop 2017. B. J. Dietvorst, et al. Algorithm Aversion: People Erroneously Avoid Algorithms After Seeing Them Err. J. Exp. Psychol. 2015. [Gupta+ '16] 9

w զʑ͸ίϛϡχέʔγϣϯʹ͓͍ͯɺϐον΍ԻྔΛมԽͤ͞ଞऀͷ஫ҙΛҾ͘ w ೴ܭଌʹΑΓɺແҙࣝతʹͦ͏ͨ͠มԽʹ൓Ԡ͍ͯ͠Δͱ֬ೝ͞Ε͍ͯΔ<9V> w ͜ͷ஌֮ಛੑΛར༻͢Δ͜ͱͰɺ׬ᘳͰͳ͍ػցֶशϞσϧΛ࢖͍ͳ͕Β΋ ஫ҙΛऔΓ໭ͨ͢Ίͷߦಈม༰ϑΟʔυόοΫΛ࣮ݱͰ͖Δ Y. Xu.
Speech Melody as Articulatorily Implemented Communicative Functions. Speech Comm. 2005. 11

.JOEMFTTDPNQVUJOH<"EBNT > A. T. Adams et al. Mindless Computing: Designing
Technologies to Subtly Influence Behavior. UbiComp 2015. U. Lyngs et al. Self-Control in Cyberspace: Applying Dual Systems Theory to a Review of Digital Self-Control Tools. ACM CHI 2019. D. Kahneman. Thinking, Fast and Slow. 2011. ΧʔωϚϯͷೋॏաఔཧ࿦ʹجͮ͘σβΠϯ <,BIOFNBO -ZOHT > • γεςϜ௚ײతͳ଎͍ࢥߟ • γεςϜ࿦ཧతͳ஗͍ࢥߟ γεςϜΛ௨ͯ͠ߦಈม༰ΛੜΈग़͢ʹ͸ Ϣʔβͷಈػ͚͕ͮඞཁʹͳͬͯ͘Δ ˠਓؒͷੜಘతͳಛੑΛ׆༻͢Δ͜ͱͰ ɹϢʔβͷ࿦ཧతͳࢥߟʹґଘͤͣʹ ɹߦಈม༰ΛੜΈग़͢Ξϓϩʔν 12

஫ҙ͕ҳΕ͍ͯΔͱݕग़͞Εͨ࣌ʹ Ի੠ͷϐον΍ԻྔΛઁಈͤ͞ΔΞϓϩʔνΛఏҊ w ಛʹԻ੠Λ׆༻͢Δ͜ͱͰɺҎԼͷར఺͕ಘΒΕΔ w ϑΟʔυόοΫͷఏ͕ࣔ͋ͬͯ΋ɺֶशͦͷ΋ͷ͕தஅ͞Εͳ͍ w Αͦݟ͍ͯͯ͠ؾ͔ͮͳ͍ͱ͍ͬͨՄೳੑ͕ͳ͘ɺௌ֮͸ಧ͖΍͍͢ w ௥Ճͷ֎෦σόΠεΛ༻͍ͳͯ͘΋ϑΟʔυόοΫΛఏࣔͰ͖Δ
13

ΦϯσϚϯυतۀΛ໛ͨ͠؀ڥͰɺεςοϓʹ෼͚ͯ༗ޮੑΛݕূ ).JOEMFTT"UUSBDUPS ͸Ϣʔβͷ໌ࣔతͳҙࣝ෇͚ͳ͘ ߦಈม༰ΛੜΈग़͢ϑΟʔυόοΫσβΠϯͱͯ͠ػೳ͢Δ ˠਓؒͷΞϊςʔλΛ׆༻࣮ͨ͠ݧઃܭͰ֬ೝ ).JOEMFTT"UUSBDUPS ͸׬ᘳͰͳ͍ػցֶशϞσϧͱ૊Έ߹Θͤͯ΋ ػೳ͢Δ͚ͩͰͳ͘ɺ ΞϥʔτͷఏࣔΑΓϢʔβͷධՁ΋ߴ͘ͳΔ ˠ࣮ࡍʹطଘͷػցֶशϞσϧΛ૊ΈࠐΜ࣮ͩݧઃܭͰ֬ೝ
14

w հೖ͋Γ৚݅Ͱ͸ɺϦϞʔτͰ؍࡯͍ͯ͠ΔΞϊςʔλʔ͕ ࢀՃऀͷूதঢ়ଶͷมԽΛ൑ఆͯ͠ϦΞϧλΠϜʹه࿥ w ूத͕ͦΕͨͱ͖ʹͷ֬཰ͰԻ੠ϑΟʔυόοΫΛఏࣔ w հೖͳ͠৚݅͸ɺೝ஌ෛՙΛൺֱ͢ΔͨΊͷϕʔεϥΠϯͱͯ͠ಋೖ 15 ࣮ݧϑΟʔυόοΫͦͷ΋ͷͷޮՌݕূ

w ϑΟʔυόοΫͷఏ͕ࣔूதΛ໭͢ͷʹཁͨ࣌ؒ͠Λ༗ҙʹ࡟ݮ͠ɺ ఏҊख๏ͷߦಈม༰ΛੜΈग़͢खஈͱͯ͠ͷ༗ޮੑΛ֬ೝ w /"4"5-9Ͱͷೝ஌ෛՙ΋มΘΒͣɺ໌ࣔతҙࣝ෇͚ͳ͠ʹػೳ͢Δͱࣔࠦ 16 ࣮ݧϑΟʔυόοΫͦͷ΋ͷͷޮՌݕূ

࣮ݧِཅੑͷϦεΫͷதͰͷ༗ޮੑ w ໊ͷࢀՃऀ͕ͭͷߨٛಈըΛࢹௌ͠ɺͦͷ಺༰Λཁ໿ w ࢀՃऀ࣮ؒݧͰɺఏҊख๏ɾ໌ࣔతͳΞϥʔτɾհೖͳ͠Λൺֱ w հೖ͸طଘͷػցֶशϞσϧͰൃಈ͞ΕΔͨΊɺ ِཅੑͷϦεΫ͕͋Δঢ়ଶ 17

ूத͍ͯ͠ͳ͍࣌ؒͷ ࡟ݮʹߩݙ ूத͕ͦΕΔස౓ࣗମ͸ มԽͳ͠ ໌ࣔతͳΞϥʔτΑΓ ମݧͷධՁ͸޲্ 18 ࣮ݧِཅੑͷϦεΫͷதͰͷ༗ޮੑ

ೋॏաఔཧ࿦ΛഎܠʹɺԻ੠ͷϐονͱԻྔΛઁಈͤ͞Δ͜ͱͰ ϢʔβͷूதΛҡ࣋͢ΔΠϯλϥΫγϣϯσβΠϯΛఏࣔ͠ɺ ػցֶशͷِཅੑʹର͢Δϩόετੑ͕͋Δ͜ͱΛݕূ .JOEMFTT"UUSBDUPSػցֶशͷِཅੑΛ ౿·͑ͨௌ֮ϑΟʔυόοΫσβΠϯ

$BU"MZTUେن໛ੜ੒ϞσϧΛ༻͍ͨ ஌తλεΫͷઌԆ͹͠๷ࢭΠϯλϥΫγϣϯ Riku Arakawa*1, Hiromu Yakura*2,3, Masataka Goto3 *: equal
contribution 1 : Carnegie Mellon University 2: Max-Planck Institute for Human Development (Previously: University of Tsukuba) 3: National Institute of Advanced Industrial Science and Technology (AIST) 20

21 S. Noy, et al. Experimental evidence on the productivity
effects of generative artificial intelligence. Science 2023. https://www.technologyreview.com/2023/07/13/1076199/ --.͸஌తੜ࢈ੑΛ ߴΊΔ͔

--. Λ͏·͘࢖͍͜ͳ͢͜ͱ͸೉͘͠ ׆༻৔໘͸ґવͱͯ͠ݶΒΕ͍ͯΔ ಛఆͷυϝΠϯ஌͕ࣝ଍Γͳ͍৔߹΍ ͜͏ͨ͠ݶքΛલఏʹ͠ͳ͕Β΋ ׆༻ൣғΛ޿͛Δ͜ͱ͸Ͱ͖ͳ͍͔ʁ ৽ͨͳΞΠσΞͷఏࣔʹ΋ϋʔυϧ͕ 22

Ծઆ--. ͷଘࡏ͸࡞ۀͷʮઌԆ͹͠บʯղফʹ໾ཱͭ Ծఆ׬ᘳͰ͸ͳ͍ੜ੒݁ՌͰ͋ͬͯ΋ Ϣʔβͷ஫ҙΛλεΫʹҾ͖໭͢ͷʹ͸࢖͑Δ طଘख๏ ਐḿͷՄࢹԽʹΑΔ ಈػ͚ͮ<-JV > αΠτϒϩοΧʔ <,PWBDT
> ఏҊख๏ G. Kovacs, et al. Rotating Online Behavior Change Interventions Increases Effectiveness But Also Increases Attrition. ACM CSCW 2018. Y. Liu, et al. Supporting Task Resumption Using Visual Feedback. ACM CSCW 2014. --. Λ༻͍ͨհೖ 23

σϞεϥΠυ࡞੒த 24

$BU"MZTU֓ཁ λεΫͷछྨʹؔΘΒͣԠ༻Ͱ͖ΔΞϓϩʔνͱͯ͠ઃܭ จॻࣥචͱεϥΠυ࡞੒ͷͭͷγφϦΦͰޮՌݕূ 26

·ͣϢʔβ͕௕࣌ؒʹ౉ͬͯ࡞ۀΛதஅ͍ͯ͠Δ৔໘Λݕ஌ $BU"MZTU֓ཁ 27

Α͋͘ΔհೖΞϓϩʔν --.Λ࢖ͬͯ ϝοηʔδͷఏࣔΛ௒͑ͨ Ξϓϩʔν͸Ͱ͖ͳ͍͔ $BU"MZTUհೖ ɾόϦΤʔγϣϯΛ૿΍͢ ɾ಺༰ΛύʔιφϥΠζ͢Δ ︙ 28

--.͕தஅͨ͠࡞ۀͷଓ͖ͷΞΠσΞΛఏࣔ͢Δ͜ͱͰ w NPUJWBUJPO࡞ۀͷ࠶։΁ͷಈػ͚ͮΛߦ͏ w BCJMJUZ࡞ۀͷ࠶։΁ͷೝ஌తϋʔυϧΛԼ͛Δ R. E. Petty, et al.
The Elaboration Likelihood Model of Persuasion. Adv. Exp. Soc. Psychol. 1986. B. J. Fogg. A Behavior Model for Persuasive Design. ACM Persuasive 2009. &MBCPSBUJPOMJLFMJIPPENPEFM &-. <1FUUZ > w ߦಈม༰ΛੜΈग़͢Ҽࢠͱͯ͠ NPUJWBUJPOͱBCJMJUZͷ྆ํ͕ඞཁͱࢦఠ <'PHH> $BU"MZTUհೖ 29

όϦΤʔγϣϯͷ͋Δհೖͱͯ͠ػೳ͢Δ͚ͩͰͳ͘ $BU"MZTUհೖ 30

গ͠త֎Εͩͬͨͱͯ͠΋ ߟ͑Δʮωλʯ͕͋Δ͜ͱͰ஫ҙΛҾ͖໭ͤΔ $BU"MZTUհೖ 31

w (15 Ͱதஅ͞Ε͍ͯΔλεΫͷଓ͖ͷΞΠσΞΛੜ੒ w εϥΠυ࡞੒࣌ʹ͸ %JGGVTJPO.PEFM Ͱը૾΋ੜ੒ $BU"MZTU࣮૷ 32

Ϣʔβ࣮ݧαΠτϒϩοΧʔͱͷൺֱ จॻࣥච εϥΠυ࡞੒ w )தஅͨ͠λεΫͷଓ͖Λఏࣔ͢Δ͜ͱͰɺ$BUBMZTU ͸Ϣʔβͷ஫ҙΛҾ͖໭ͤΔ w )Ϣʔβͷ஫ҙΛҾ͖໭͢͜ͱͰɺ$BUBMZTU ͸࡞ۀͷ࠶։ΛޮՌతʹଅͤΔ w
)$BU"MZTU Λ༻͍Δ͜ͱͰɺλεΫͷதஅΛճආ͠ੜ࢈ੑΛߴΊΒΕΔ w )$BU"MZTU͸Ϣʔβͷೝ஌ෛՙΛߴΊͣʹ࢖͏͜ͱ͕Ͱ͖Δ 33

w *HOPSBODFSBUF௨஌͕ແࢹ͞Εׂͨ߹ w *OUFSFTUSFUSJFWBMUJNF௨஌͔Β࡞ۀ࠶։·Ͱͷॴཁ࣌ؒ w 1SPHSFTTBGUFSSFTVNQUJPO࠶։͔ΒҰఆ࣌ؒͷ࡞ۀྔʢจࣈ਺౳ʣ w 5PUBMUJNF༩͑ΒΕͨλεΫશମͷॴཁ࣌ؒ w 4VCKFDUJWFRVBMJUZΫϥ΢υϫʔΧʹΑΔ੒Ռ෺΁ͷΫΦϦςΟධՁ
w $PHOJUJWFMPBE/"4"5-9 Ͱܭଌ͞Εͨೝ஌ෛՙ w 4ZTUFNVTBCJMJUZ464 Ͱܭଌ͞ΕͨϢʔβϏϦςΟධՁ ? ? શମͷॴཁ࣌ؒͱୈࡾऀͷΫΦϦςΟධՁ͸͕ࠩͳ͔ͬͨ΋ͷͷ $BU"MZTU͸શମతʹ༗ҙͳޮՌΛࣔͨ͠ Ϣʔβ࣮ݧࢦඪٴͼ݁Ռ จॻࣥච εϥΠυ࡞੒ 34

w $BU"MZTU Λ೔ؒʹΘͨͬͯఏڙ w λεΫ౳͸༩͑ͣࣗ༝ʹ࢖༻ w ࢖༻ޙʹ൒ߏ଄ԽΠϯλϏϡʔ Ϣʔβ࣮ݧ௕ظతͳ࢖༻࣮ݧ 35

w ߦಈʹ༩͑ͨӨڹ w "*ͷਫ਼౓ʹ͍ͭͯͷײ૝ w $BU"MZTUͷ໾ׂ w 3FNJOEFS*EFBUPS1FFS w վળͷ༨஍
ΠϯλϏϡʔ݁Ռ ࢖༻ঢ়گ w ೔ؒͷؒͰ࢖༻ස౓ʹࠩ͸ݟΒΕͣ த௕ظతͳޮՌ΍ϢʔβϏϦςΟΛࣔࠦ ৄࡉ͸࿦จΛࢀর͍ͩ͘͞ʂ Ϣʔβ࣮ݧ௕ظతͳ࢖༻࣮ݧ 36

ᶃ--. Λ fi OFUVOJOH ͢΂͖͔ υϝΠϯ΍λεΫ͝ͱʹϞσϧΛνϡʔχϯά͢Ε͹ ʮઌԆ͹͠ʯճආ͚ͩͰͳ͘ɺ௚઀తͳ࡞ۀͷޮ཰Խʹ΋ͭͳ͕Δ طଘͷϞσϧΛͦͷ··࢖͑͹௥Ճίετͳ͠ʹ ෯޿͍υϝΠϯͷ࡞ۀΛࢧԉͰ͖Δ ٞ࿦
37

ᶄ"*ͷೳಈతͳ࢖༻ͱडಈతͳ࢖༻ ʮੜ੒ϘλϯʯʹΑΔೳಈతͳ࢖༻ߴ͍ظ଴஋㲗ࣦ๬ͷϦεΫ ʮ$BU"MZTUʯͷΑ͏ͳडಈతͳ࢖༻௿͍ظ଴஋㲗ϦεΫͷ௿ݮ 6OSFNBSLBCMF"*<:BOH > ໨ཱͨͣʹ"*͕Ձ஋Λൃش͢ΔΠϯλϥΫγϣϯσβΠϯ ≒ Q. Yang, et
al. Unremarkable AI: Fitting Intelligent Decision Support into Critical, Clinical Decision-Making Processes. ACM CHI 2019. ٞ࿦ 38

Writing Slide-Editing ϑΝΠϯνϡʔχϯάʹΑͬͯਫ਼౓ΛߴΊɺ௚઀λεΫΛޮ཰Խ͢ΔͷͰ͸ͳ͘ ׬ᘳͰ͸ͳ͘ͱ΋طଘͷϞσϧΛ͏·͘׆༻͢Δ͜ͱͰ Ϣʔβͷߦಈม༰Λಋ͖ɺ%JHJUBM8FMMCFJOHͷ޲্Λ໨ࢦ͢ΠϯλϥΫγϣϯ Composition ɾɾɾ $BU"MZTUେن໛ੜ੒ϞσϧΛ༻͍ͨ ஌తλεΫͷઌԆ͹͠๷ࢭΠϯλϥΫγϣϯ 39

ػցֶशͰώτͷߦಈΛม͑Δ ػցֶश͕ώτͷߦಈΛม͑Δ 40

&NQJSJDBMFWJEFODFPG-BSHF-BOHVBHF.PEFMT JOGMVFODFPOIVNBOTQPLFODPNNVOJDBUJPO 41 Hiromu Yakura*, Ezequiel Lopez-Lopez*, Levin Brinkmann*, Ignacio
Serna, Prateek Gupta, Iyad Rahwan *: equal contribution Max-Planck Institute for Human Development

• $IBU(15 ͸ͳ͔ͥ EFMWF ͱ͍͏ ୯ޠΛ࢖͍͕ͪͱ͍͏ݱ৅͕໌Β͔ʹ • ଞʹ΋ $IBU(15 ʹಛ௃తͳ୯ޠ͕
ൃݟ͞ΕɺͦΕΒͷՊֶ࿦จͰͷ ग़ݱස౓ͷٸ૿΋ࢦఠ͞Ε͍ͯΔ $IBU(15 ͷ EFMWF όΠΞε 42 W. Liang, et al. Mapping the Increasing Use of LLMs in Scientific Papers. Proc. CoLM (2024). https://pshapira.net/2024/03/31/delving-into-delve/

ࣗ෼Ͱ΋ৼΓฦͬͯΈΔͱ • ࣗ਎΋Α͘ $IBU(15 ʹӳޠϝʔϧͷจষͳͲΛఴ࡟ͯ͠΋Β͏ • ͦͯ͠ɺ$IBU(15 Λ࢖͍ͬͯͳ͍ͱ͖ʹ΋ EFMWF ͱ͍͏୯ޠΛ࢖͏บ͕
఻છ͍ͯ͠Δؾ͕͢Δ ࣮ࡍʹ$IBU(15͕զʑͷίϛϡχέʔγϣϯʹ ͲΜͳมԽΛ΋ͨΒ͍ͯ͠Δͷ͔Λݕূ 43

• ੈքͷֶज़ػؔͷ :PV5VCF νϟϯωϧ͔Β ෼Ҏ্ͷಈըສ݅Λऩू͠ॻ͖ى͜͠ ◦ ӳޠ฼ޠ࿩ऀΑΓୈೋݴޠ࿩ऀͷ΄͏͕ ӨڹΛड͚ΔͷͰ͸ͳ͍͔ͱ͍͏Ծఆ ◦ ΞΧσϛΞ͸͢Ͱʹࣥච౳Ͱ$IBU(15Λ
େن໛ʹར༻͍ͯ͠Δ͜ͱ͕໌Β͔ ◦ ௕࣌ؒͷϓϨθϯ౳Λ͢΂ͯ$IBU(15Ͱ ݪߘ࡞੒͢ΔՄೳੑ͸௿͘ɺ ൃݴ͸TQPOUBOFPVTͳ಺༰ͱΈͳͤΔ Active institutions on ROR (n=20,662) Institutions with a channel identified (n=16,667) Institution with no channel identified (n=3,995) Videos from the channels (n=2,958,103) Videos with a English title (n=1,613,839) Videos with a non-English title (n=1,344,264) Videos longer than 20 minutes (n=364,916) Videos shorter than 20 minutes (n=1,248,913) Global institutions Videos shorter than 99 percentile (n=361,297) Videos longer than 99 percentile (n=3,619) σʔλऩू 44

• ҎԼͷܗࣜͰ $IBU(15 ެ։ޙͷ มԽྔΛϞσϦϯά • ݁Ռɺ(15 ʹಛ༗ͱ͞ΕΔ ୯ޠ͸ݦஶͳ૿ՃΛݟͤͨ EFMWF͸Ի੠ίϛϡχέʔγϣϯͰ΋૿͍͑ͯͨ
46

ςΩετͰͷ෼ੳͷ݁Ռͱͷؔ࿈ੑ΋֬ೝ • ͞Βʹطଘݚڀ͕࿦จͷղੳ͔Β ࢉग़ͨ͠୯ޠͷ (15 Β͠͞ͱ ૬ؔ͢Δ܏޲΋֬ೝ • ࣥචͳͲͷλεΫΛ௨ͯ͠ ਓ͕ؒӨڹ͞Εͨͱ͍͏ܦ࿏Λࣔࠦ
47

ΞϑϦΧӳޠ͕ $IBU(15 ܦ༝Ͱશੈքʹ఻೻͍ͯ͠Δʁ • ͔͠΋͜͏ͨ͠୯ޠ͸έχΞӳޠͳͲͰ Α͘࢖ΘΕ͍ͯΔ͜ͱ΋֬ೝ • 0QFO"* ͕ΞϊςʔγϣϯΛ҆Ձʹ֎஫ͨ͠ ݁Ռ͕શੈքʹӨڹ͍ͯ͠Δͱ͍͏Մೳੑ΋
48 https://time.com/6247678/openai-chatgpt-kenya-workers/

ώτͱ"*ͷ૬ޓతͳߦಈม༰ͷՄೳੑ "*ʹΑΔ ߦಈม༰ όΠΞε͞Εͨ σʔλͷ஝ੵ ࣍ੈ୅Ϟσϧͷ ৽σʔλͰͷֶश
ࣾձ΁ͷ Өڹͷྦྷੵ 👩 🤖 📝 ݴޠදݱ ௌऔߦಈ طଘͷ σʔληοτ ࠓޙͷ σʔληοτ ︙ 📝 📝🤖 🎧 🎧🤖 ︙ 📝🤖 🎧🤖 🤖 🤖 🤖 ✨ ✨ ✨ ︙ 🤖 🤖 🤖 ✨ ✨ ✨ ︙ όΠΞεΛ Ϟσϧ͕ ڧԽ "* ٕज़ͷ΋ͨΒ͔ᷮ͢ͳมԽ͕ɺҙਤͤ͵ํ޲΁ͱ ਓྨશମͷදݱ΍ߦಈΛ༠ಋ͍ͯ͘͠Մೳੑ͕͋Δ 👱 🤖 👩 🤖 📝 ✨ 🤖 ✨ 👱 🤖✨ 🤖 ✨ 49

͓·͚ 50

"* Λ 5SJDLPS5SFBU ͢ΔήʔϜ "* ͷٖਓԽ͞Εͨײ৘දݱ͕ ਓؒͷߦಈʹ༩͑ΔӨڹΛௐࠪ TQPPLUIFNBDIJOFDPN

機械学習でヒトの行動を変える

機械学習でヒトの行動を変える

More Decks by Hiromu Yakura

Other Decks in Research

Featured

Transcript