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

自然言語処理の基礎と応用 〜 料理と医療を題材として 〜 /JADI2021

558ffd4e5cf6c95e60c91333677b93c4?s=47 j.harashima
September 10, 2021
0
170

自然言語処理の基礎と応用 〜 料理と医療を題材として 〜 /JADI2021

558ffd4e5cf6c95e60c91333677b93c4?s=128

j.harashima

September 10, 2021
Tweet

More Decks by j.harashima

See All by j.harashima
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
1
2.9k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
0
8k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
0
2.3k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
0
54
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
0
4.9k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
1
10k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
0
3.6k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
0
5.5k
558ffd4e5cf6c95e60c91333677b93c4?s=48 junharashima
1
59

Featured

See All Featured
91cefdf141106fa10674aae74ce05890?s=48 yeseniaperezcruz
307
32k
282d599b414022eba5096510f675b6e2?s=48 chrislema
231
16k
4262b9cfacfb6b2c77f23e1d0310cd3e?s=48 myddelton
110
11k
C7393b7ba7ec9c8890dd77d209fbb3c9?s=48 maltzj
504
37k
Eb8975af8e49e19e3dd6b6b84a542e26?s=48 qrush
288
19k
39488f9d172ab92fd352f2cd7b73258d?s=48 mza
82
4.4k
5f2da528927a2ec9ba4fec2069cbc958?s=48 kneath
298
39k
0bce06bd06ee7a2c4f5500f25dcf7f18?s=48 paulrobertlloyd
72
3.9k
A9a491b0fcbe0fbce3d64063a37add99?s=48 rmw
15
1k
E9221b172e78e888355fa2fe4541b595?s=48 mclloyd
PRO
6
7.5k
9128d500301ae51524e887bb680f471d?s=48 caitiem20
314
18k
F716e5f737fcfb03f2cf8d1a184f1dbe?s=48 nonsquared
84
3.6k

Transcript

  1. ࣗવݴޠॲཧ ͷجૅͱԠ༻ ʙྉཧͱҩྍΛ୊ࡐͱͯ͠ʙ ݪౡ७

  2.   ژ౎େֶʹͯࣗવݴޠॲཧͷݚڀࣨʹ ഑ଐ ത࢜ʢ৘ใֶʣΛऔಘɺΫοΫύου ʹೖࣾ ݪౡ७  ݚڀ։ൃ෦Λ৽ઃ 

    ϨγϐݕࡧνʔϜΛ৽ઃɺϦʔμʔʹ ब೚ʢݚڀ։ൃ෦෦௕΋݉຿ʣ  ಉ෦෦௕ʹब೚

  3. ໔੹ࣄ߲ ࠓ೔ͷൃද͸ʮΩονϯɾΠϯϑΥϚςΟΫεʯͷ ಺༰Λϕʔεʹ͍ͯ͠·͢ɻ ͜ͷຊ͸ɺΫοΫύουͷࣾһͱͯ͠Ͱͳ͘ɺ ݸਓͱͯ͠ॻ͍ͨ΋ͷͰ͢ɻ ൃදʹؔ͢Δ࣭͝໰౳͸ɺձࣾʹͰͳ͘ɺ ݪౡݸਓʹ͓د͍ͤͩ͘͞ɻ

  4. ໨࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ

  5. ࣗવݴޠॲཧͱ͸ʁ ࣗવݴޠʢʹݴ༿ʣΛίϯϐϡʔλͰѻ͏ͨΊͷٕज़ ࣗવݴޠࣗવൃੜతʹੜ·ΕͨݴޠʢFH ೔ຊޠʣ ਓ޻ݴޠਓ޻తʹ࡞ΒΕͨݴޠʢFH $ݴޠʣ

  6. ৘ใݕࡧʜϢʔβ͕ٻΊΔ৘ใΛจॻू߹͔Βݕࡧ ػց຋༁ʜҙຯΛอͬͨ··ɺ͋ΔݴޠͷจॻΛผͷݴޠͷจॻʹม׵ จॻ෼ྨʜ༧ΊఆΊΒΕͨೋͭҎ্ͷΫϥεͷ͍ͣΕ͔ʢPS͍͔ͭ͘ʣʹจॻΛ෼ྨ จॻཁ໿ʜ৘ใΛอͬͨ··ɺҰͭʢPSෳ਺ʣͷจॻΛ୹ॖ ৘ใநग़ʜจॻ͔Βॏཁͳ৘ใΛநग़ FUD Ԡ༻ྫ ৘ใݕࡧͷҰྫ

  7. ػցֶशͱ͸ʁ ɾσʔλʹજΉύλʔϯΛػցతʹݟ͚ͭΔͨΊͷٕज़ ػցֶशʢಛʹχϡʔϥϧωοτϫʔΫʣʹΑΔ༷ʑͳൃల ɾ8PSE7FDʹΑΔҙຯͷϕΫτϧԽʢ.JLPMPWFUBM ʣ ɾ#&35ʹΑΔ֤छݴޠཧղλεΫͷੑೳվળʢ%FWMJOFUBM ʣ ۙ೥ɺଟ͘ͷاۀ͕ࣗવݴޠॲཧʹ஫໨ ػցֶशͷ୆಄ 

       ε ϙ ϯ α ồ ਺   ਓ޻஌ೳֶձશࠃେձ      ը૾ͷೝࣝɾཧղγϯϙδ΢Ϝ ݴޠॲཧֶձ೥࣍େձ ˞೥ʹεϙϯαʔ਺͕ݮ͍ͬͯΔͷ͸ίϩφͷӨڹ

  8. ೔ຊޠʹ͓͚Δࣗવݴޠॲཧ ୯ޠ୯Ґ จઅ୯Ґ จ୯Ґ Ԡ༻ʢFH ৘ใݕࡧɺػց຋༁ʣ ߏจղੳ ܗଶૉղੳɹݻ༗දݱೝࣝ ड़ޠ߲ߏ଄ղੳɹڞࢀরղੳ ҙຯ୯Ґ

    ಉٛؔ܎ೝࣝ ؚҙؔ܎ೝࣝ ຊൃදͰ͸੺ࣈͷٕज़ʹ͍ͭͯͷΈ঺հʢͦͷଞʹ͍ͭͯ͸੿ஶΛ͝ཡ͍ͩ͘͞ʣ

  9. ܗଶૉղੳ จதͷܗଶૉʢ㲈୯ޠʣͷڥքͱ඼ࢺɺݪܗΛೝࣝ͢Δॲཧ ೔ຊޠ͸୯ޠ͕εϖʔεͰ۠੾ΒΕ͍ͯͳ͍ˠܗଶૉղੳ͸೔ຊޠॲཧͷୈҰา ϑϥϯε ଠ࿠ ͸ Λ ৯΂ ύϯ c

    ໊ࢺ ໊ࢺ ಈࢺ ॿࢺ ໊ࢺ c c c ৯΂Δ ଠ࿠ ͸ Λ ϑϥϯε ύϯ ݪܗ ڥք ඼ࢺ c ॿࢺ ଠ࿠͸ϑϥϯεύϯΛ৯΂ͨ ೖྗ ग़ྗ ͨ ͨ c ॿಈࢺ จ

  10. ݻ༗දݱೝࣝ จதͷݻ༗දݱʢFH ਓ໊ɺ஍໊ɺʜʣΛೝࣝ͢Δॲཧ ܗଶૉղੳͷग़ྗΛೖྗͱ͠ɺݻ༗දݱͷϥϕϧΛग़ྗ ϑϥϯε ଠ࿠ ͸ Λ ৯΂ ύϯ

    c ໊ࢺ ໊ࢺ ಈࢺ ॿࢺ ໊ࢺ c c c ৯΂Δ ଠ࿠ ͸ Λ ϑϥϯε ύϯ ݪܗ ڥք ඼ࢺ c ॿࢺ ೖྗ ͨ ͨ c ॿಈࢺ ग़ྗ ਓ໊ ͦͷଞ ͦͷଞ ͦͷଞ ͦͷଞ ͦͷଞ ஍໊ ϥϕϧ

  11. ಉٛؔ܎ೝࣝ ͋Δޠ۟ͱผͷޠ͕۟ಉ͡ҙຯ͔ҧ͏ҙຯ͔Λೝࣝ͢Δॲཧ ୯ޠಉ࢜Ͱ͸ಉ͡ҧ͏ҙຯͰ΋ɺผͷ୯ޠͱ߹ΘͤΔͱҧ͏ಉ͡ҙຯʹͳΔ͜ͱ΋ ৯΂Δ ͨ΂Δ ৯΂Δ ৯͏ ৯΂Δ ҿΉ ୯ޠϖΞ

    ೖྗ ग़ྗ ೝࣝ݁Ռ ಉ͡ ҧ͏ ಉ͡ ৯͏ ࢖͏ ҧ͏ ϑϨʔζϖΞ ೝࣝ݁Ռ ϝϞϦΛ৯͏ ϝϞϦΛ࢖͏ ಉ͡ Ұഋ৯͏ Ұഋ৯΂Δ ҧ͏ ʢͩ·͞ΕΔʣ ೖྗ ग़ྗ

  12. ࣗવݴޠॲཧͷπʔϧɾσʔλ .F$BCʢܗଶૉղੳثʣ ೔ຊޠϫʔυωοτʢγιʔϥεʣ γιʔϥεʢྨޠࣙయʣ ෼ྨޠኮදɺ೔ຊޠޠኮେܥɺʜ ੜίʔύε ੨ۭจݿɺ8JLJQFEJBɺʜ ஫ऍ෇͖ίʔύε ژ౎େֶςΩετίʔύεɺ/"*455FYU$PSQVTɺʜ ʜ

    ܗଶૉղੳث 4VEBDIJɺ+6."/ʢ+6."/ ʣɺʜ ߏจղੳث ,/1ɺ$BCP$IBɺʜ ػցֶशϥΠϒϥϦ 1Z5PSDIɺ5FOTPS'MPXɺʜ ʜ

  13. ໨࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ

  14. ༷ʑͳՄೳੑ Ϩετϥϯ Ϩγϐ ϝχϡʔ ޱίϛ ৘ใݕࡧ ػց຋༁ จॻ෼ྨ ʜ ྉཧυϝΠϯʹ͓͚ΔԠ༻

    ʹ º ྉཧυϝΠϯʹ ͓͚ΔςΩετ ࣗવݴޠॲཧͷ Ԡ༻ ʜ ৘ใநग़ ʢͷ঺հϖʔδʣ

  15. Ϩγϐº৘ใݕࡧ ͨ·Ͷ͗ πφͱͨ·Ͷ͗ͷ ࿨෩ύελ ߜΓࠐΉ   ඼ ۄೢͱੜᇙͷ ͔͋ͬͨຯ઻ो

    ΧϦΧϦ̇ ΦχΦϯϦϯά ۄͶ͗ͱτϚτͷ αϥμ ΦχΦϯάϥλϯ εʔϓ  ܗଶૉղੳ΍ݻ༗දݱೝࣝΛ֤Ϩγϐʹద༻͠ɺ 
 Ϩγϐதͷ୯ޠ΍ݻ༗දݱΛೝࣝ 
  ֤୯ޠʹ͍ͭͯಉٛޠΛల։ 
 FH ͨ·Ͷ͗ʹۄೢʹۄͶ͗ʹΦχΦϯ 
  ΫΤϦʢ͓Αͼɺͦͷಉٛޠʣ͕ग़ݱ͢ΔϨγϐΛ 
 ݕࡧ݁Ռͱͯ͠ఏࣔ

  16. Ϩγϐºจॻ෼ྨ ͨ·Ͷ͗ ओࡊ ओ৯ ۄೢͱੜᇙͷ ͔͋ͬͨຯ઻ो ෭ࡊ ो෺ ߜΓࠐΉ 

     ඼ ΦχΦϯάϥλϯ εʔϓ ཛͱ౾෗ͱۄೢͷ ༏͍͠εʔϓ ͨ·Ͷ͗ͱਓࢀͱ Ωϟϕπͷϙτϑ  ܗଶૉղੳ΍ݻ༗දݱೝࣝΛ֤Ϩγϐʹద༻͠ɺ 
 Ϩγϐதͷ୯ޠ΍ݻ༗දݱΛೝࣝ 
  ֤୯ޠʹ͍ͭͯಉٛޠΛల։ 
 FH ͨ·Ͷ͗ʹۄೢʹۄͶ͗ʹΦχΦϯ 
  ֤ΫϥεʢFH ो෺ʣʹॴଐ͢΂͖ϨγϐΛਓखͰ 
 ऩू͠ɺͦΕΒʹग़ݱ͢Δ୯ޠͷύλʔϯΛ೺Ѳ 
  Ͱऩूͨ͠ϨγϐҎ֎ͷϨγϐʹύλʔϯΛద༻͠ɺ 
 ֤Ϋϥεʹ෼ྨ

  17. y1 yJ y2 … canonical form of the ingredient decoder

    original string of an ingredient encoder x1 x2 xI … softmax softmax softmax … Ϩγϐதͷදه ৯඼੒෼දதͷදه දه༳Ε ਓࢀ ʹΜ͡Μ ུه ৽ͨ· ৽ͨ·Ͷ͗ εϖϧϛε ͋΅͕Ͳ ΞϘΧυ ͦͷଞ അླḕ ͡Ό͕͍΋ dense amount encoder xa1 xa2 xaJ … title encoder xt1 xt2 xtI … softmax y serving λΠτϧ .ࡼ෼౳ /ਓ෼ ίίφοπΦΠϧೖΓ΄͏ΕΜ૲ΧϨʔ ࡼ෼  ͜ΜͿɾͻ͖͡ɾ΄ͨͯ͝͸Μ ߹෼  άϥϊʔϥΫοΩʔ ຕ  ύϯϓΩϯλϧτ DNλϧτܕ  ৄ͘͠͸ɺJun Harashima, Makoto Hiramatsu, and Satoshi Sanjo. Calorie Estimation in a Real-World Recipe Service. In Proceedings of the 32nd Annual Conference on Innovative Applications of Arti fi cial Intelligence (IAAI-20), 2020. Λ͝ཡ͍ͩ͘͞ɻ Ϩγϐºػց຋༁ʢΧϩϦʔܭࢉʣ  Ϩγϐதͷ֤ࡐྉͷදهΛ৯඼੒෼දதͷදهʹม׵  ʮ.ࡼ෼ʯ΍ʮ.߹෼ʯɺʮ.ຕʯͳͲΛʮ/ਓ෼ʯʹม׵  ֤ࡐྉͷΧϩϦʔΛ৯඼੒෼ද͔Βऔಘ͠ɺ߹ܭ  ߹ܭΛ/Ͱׂͬͯɺਓ෼ͷΧϩϦʔΛऔಘ ˞ผݴޠ΁ͷ຋༁ʢҰൠతͳػց຋༁ʣͰ͸ͳ͘ɺಉҰݴޠ΁ͷ຋༁ʢݴ͍׵͑ʣ

  18. ໨࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ

  19. ༷ʑͳՄೳੑ Χϧς ࿦จ Ϩηϓτ αϚϦ ৘ใݕࡧ ػց຋༁ จॻ෼ྨ ʜ ҩྍυϝΠϯʹ͓͚ΔԠ༻

    ʹ º ҩྍυϝΠϯʹ ͓͚ΔςΩετ ࣗવݴޠॲཧͷ Ԡ༻ ʜ ৘ใநग़

  20. ࿦จº৘ใݕࡧ Ϩγϐº৘ใݕࡧͱಉ͡ϩδοΫʢͷ͸ͣʣ

  21. Χϧςº৘ใநग़ʢ͍ΘΏΔɺϏοάσʔλͷ׆༻ʣ ೔ܦ95&$)ʮΧϧςղੳͰਫ਼ਆ࣬ױױऀͷ༧ޙΛ༧ଌʯ IUUQTYUFDIOJLLFJDPNENBUDMGFBUVSF

  22. ҩྍºࣗવݴޠॲཧͷ೉͠͞ σʔλͷॴࡏ͕Θ͔Βͳ͍ ɾ·ͩ஝ੵ͞Εͯͳ͍ʁ ɾ͞Εͯͯ΋ɺར༻ͮ͠Β͍ʁ ɾσʔλ͕ͳ͍ͱɺࣗવݴޠॲཧ͸ແྗ ɹɹɹɹσʔλͷ஝ੵɾڞ༗͕ෆՄܽ σʔλͷ಺༰͕Θ͔Βͳ͍ ɾΠϯεϦϯʁϔϞάϩϏϯʁܕʁܕʁ ɾࣗવݴޠॲཧʢFH ৘ใநग़ʣͷ݁Ռ͕ղऍෆೳ

    ɾղऍ͕ෆೳͩͱɺվળ΋ෆೳ ɹɹɹɹҩྍͷϓϩͱͷ࿈ܞ͕ෆՄܽ ҩྍͷૉਓ͔Βݟͨ

  23. ໨࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ

  24. ࣗવݴޠॲཧͷجૅͱԠ༻ʙྉཧͱҩྍΛ୊ࡐͱͯ͠ʙ ࣗવݴޠॲཧͱ͸ʁ ɾࣗવݴޠʢʹݴ༿ʣΛίϯϐϡʔλͰѻ͏ͨΊͷٕज़ ɾ͜͜਺೥Ͱٸܹʹൃలɺଟ͘ͷاۀ͕஫໨ ςΩετσʔλ͑͋͞Ε͹ɺ༷ʑͳԠ༻͕ߟ͑ΒΕΔ ɾྉཧυϝΠϯʹ͓͚ΔςΩετºࣗવݴޠॲཧͷԠ༻ྉཧυϝΠϯʹ͓͚ΔԠ༻ ɾҩྍυϝΠϯʹ͓͚ΔςΩετºࣗવݴޠॲཧͷԠ༻ҩྍυϝΠϯʹ͓͚ΔԠ༻ ҩྍºࣗવݴޠॲཧͷ೉͠͞ ɾσʔλͷॴࡏ͕Θ͔Βͳ͍ˠσʔλͷ஝ੵɾڞ༗͕ෆՄܽ ɾσʔλͷ಺༰͕Θ͔Βͳ͍ˠҩྍͷϓϩͱͷ࿈ܞ͕ෆՄܽ