Upgrade to Pro
— share decks privately, control downloads, hide ads and more …
Speaker Deck
Sign up for free
Menu
Search
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Features
All features
Private URLs
Password Protection
Custom URLS
Scheduled publishing
Remove Branding
Restrict embedding
Deck Collections
Notes
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Explore
Featured decks
Featured speakers
Programming
Technology
Storyboards
Pricing
Search
Sign in
Sign up for free
自然言語処理の基礎と応用 〜 料理と医療を題材として 〜 /JADI2021
Search
j.harashima
September 10, 2021
5.8k
0
Share
Embed
Copy iframe code
Copy JS code
Copy link
Start on current slide
自然言語処理の基礎と応用 〜 料理と医療を題材として 〜 /JADI2021
j.harashima
September 10, 2021
More Decks by j.harashima
See All by j.harashima
日本語レシピデータセットの継続的な構築と複合的な利用/JED2022
junharashima
1
15k
企業での研究開発の楽しさと苦労/WAP-Tech-Talk
junharashima
1
12k
クックパッドにおける研究開発/HCG2020
junharashima
0
5.1k
Calorie Estimation in a Real-World Recipe Service/iaai-20
junharashima
0
16k
クックパッドと機械学習(短縮版)/MLTBP
junharashima
0
11k
クックパッドと NLP/CV/nlpaper-challenge
junharashima
0
160
AI 部門の実戦投入/repro-tech-meetup
junharashima
0
5.1k
クックパッドと機械学習/uec-aix-seminar
junharashima
0
13k
実データにふれる自然言語処理インターンシップ/internship2018
junharashima
1
16k
Featured
See All Featured
Understanding Cognitive Biases in Performance Measurement
bluesmoon
32
3k
Game over? The fight for quality and originality in the time of robots
wayneb77
1
290
Refactoring Trust on Your Teams (GOTO; Chicago 2020)
rmw
35
3.8k
A Guide to Academic Writing Using Generative AI - A Workshop
ks91
PRO
1
490
Getting science done with accelerated Python computing platforms
jacobtomlinson
2
490
The Invisible Side of Design
smashingmag
301
52k
BBQ
matthewcrist
89
10k
Ten Tips & Tricks for a 🌱 transition
stuffmc
1
240
The browser strikes back
jonoalderson
0
1.7k
jQuery: Nuts, Bolts and Bling
dougneiner
66
8.6k
AI Search: Where Are We & What Can We Do About It?
aleyda
0
8k
How People are Using Generative and Agentic AI to Supercharge Their Products, Projects, Services and Value Streams Today
helenjbeal
1
340
Transcript
ࣗવݴޠॲཧ ͷجૅͱԠ༻ ʙྉཧͱҩྍΛࡐͱͯ͠ʙ ݪౡ७
ژେֶʹͯࣗવݴޠॲཧͷݚڀࣨʹ ଐ ത࢜ʢใֶʣΛऔಘɺΫοΫύου ʹೖࣾ ݪౡ७ ݚڀ։ൃ෦Λ৽ઃ
ϨγϐݕࡧνʔϜΛ৽ઃɺϦʔμʔʹ बʢݚڀ։ൃ෦෦݉ʣ ಉ෦෦ʹब
໔ࣄ߲ ࠓͷൃදʮΩονϯɾΠϯϑΥϚςΟΫεʯͷ ༰Λϕʔεʹ͍ͯ͠·͢ɻ ͜ͷຊɺΫοΫύουͷࣾһͱͯ͠Ͱͳ͘ɺ ݸਓͱͯ͠ॻ͍ͨͷͰ͢ɻ ൃදʹؔ͢Δ࣭͝ɺձࣾʹͰͳ͘ɺ ݪౡݸਓʹ͓د͍ͤͩ͘͞ɻ
࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ
ࣗવݴޠॲཧͱʁ ࣗવݴޠʢʹݴ༿ʣΛίϯϐϡʔλͰѻ͏ͨΊͷٕज़ ࣗવݴޠࣗવൃੜతʹੜ·ΕͨݴޠʢFH ຊޠʣ ਓݴޠਓతʹ࡞ΒΕͨݴޠʢFH $ݴޠʣ
ใݕࡧʜϢʔβ͕ٻΊΔใΛจॻू߹͔Βݕࡧ ػց༁ʜҙຯΛอͬͨ··ɺ͋ΔݴޠͷจॻΛผͷݴޠͷจॻʹม จॻྨʜ༧ΊఆΊΒΕͨೋͭҎ্ͷΫϥεͷ͍ͣΕ͔ʢPS͍͔ͭ͘ʣʹจॻΛྨ จॻཁʜใΛอͬͨ··ɺҰͭʢPSෳʣͷจॻΛॖ ใநग़ʜจॻ͔ΒॏཁͳใΛநग़ FUD Ԡ༻ྫ ใݕࡧͷҰྫ
ػցֶशͱʁ ɾσʔλʹજΉύλʔϯΛػցతʹݟ͚ͭΔͨΊͷٕज़ ػցֶशʢಛʹχϡʔϥϧωοτϫʔΫʣʹΑΔ༷ʑͳൃల ɾ8PSE7FDʹΑΔҙຯͷϕΫτϧԽʢ.JLPMPWFUBM ʣ ɾ#&35ʹΑΔ֤छݴޠཧղλεΫͷੑೳվળʢ%FWMJOFUBM ʣ ۙɺଟ͘ͷاۀ͕ࣗવݴޠॲཧʹ ػցֶशͷ಄
ε ϙ ϯ α ồ ਓೳֶձશࠃେձ ը૾ͷೝࣝɾཧղγϯϙδϜ ݴޠॲཧֶձ࣍େձ ˞ʹεϙϯαʔ͕ݮ͍ͬͯΔͷίϩφͷӨڹ
ຊޠʹ͓͚Δࣗવݴޠॲཧ ୯ޠ୯Ґ จઅ୯Ґ จ୯Ґ Ԡ༻ʢFH ใݕࡧɺػց༁ʣ ߏจղੳ ܗଶૉղੳɹݻ༗දݱೝࣝ ड़ޠ߲ߏղੳɹڞࢀরղੳ ҙຯ୯Ґ
ಉٛؔೝࣝ ؚҙؔೝࣝ ຊൃදͰࣈͷٕज़ʹ͍ͭͯͷΈհʢͦͷଞʹ͍ͭͯஶΛ͝ཡ͍ͩ͘͞ʣ
ܗଶૉղੳ จதͷܗଶૉʢ㲈୯ޠʣͷڥքͱࢺɺݪܗΛೝࣝ͢Δॲཧ ຊޠ୯ޠ͕εϖʔεͰ۠ΒΕ͍ͯͳ͍ˠܗଶૉղੳຊޠॲཧͷୈҰา ϑϥϯε ଠ Λ ৯ ύϯ c
໊ࢺ ໊ࢺ ಈࢺ ॿࢺ ໊ࢺ c c c ৯Δ ଠ Λ ϑϥϯε ύϯ ݪܗ ڥք ࢺ c ॿࢺ ଠϑϥϯεύϯΛ৯ͨ ೖྗ ग़ྗ ͨ ͨ c ॿಈࢺ จ
ݻ༗දݱೝࣝ จதͷݻ༗දݱʢFH ਓ໊ɺ໊ɺʜʣΛೝࣝ͢Δॲཧ ܗଶૉղੳͷग़ྗΛೖྗͱ͠ɺݻ༗දݱͷϥϕϧΛग़ྗ ϑϥϯε ଠ Λ ৯ ύϯ
c ໊ࢺ ໊ࢺ ಈࢺ ॿࢺ ໊ࢺ c c c ৯Δ ଠ Λ ϑϥϯε ύϯ ݪܗ ڥք ࢺ c ॿࢺ ೖྗ ͨ ͨ c ॿಈࢺ ग़ྗ ਓ໊ ͦͷଞ ͦͷଞ ͦͷଞ ͦͷଞ ͦͷଞ ໊ ϥϕϧ
ಉٛؔೝࣝ ͋Δޠ۟ͱผͷޠ͕۟ಉ͡ҙຯ͔ҧ͏ҙຯ͔Λೝࣝ͢Δॲཧ ୯ޠಉ࢜Ͱಉ͡ҧ͏ҙຯͰɺผͷ୯ޠͱ߹ΘͤΔͱҧ͏ಉ͡ҙຯʹͳΔ͜ͱ ৯Δ ͨΔ ৯Δ ৯͏ ৯Δ ҿΉ ୯ޠϖΞ
ೖྗ ग़ྗ ೝࣝ݁Ռ ಉ͡ ҧ͏ ಉ͡ ৯͏ ͏ ҧ͏ ϑϨʔζϖΞ ೝࣝ݁Ռ ϝϞϦΛ৯͏ ϝϞϦΛ͏ ಉ͡ Ұഋ৯͏ Ұഋ৯Δ ҧ͏ ʢͩ·͞ΕΔʣ ೖྗ ग़ྗ
ࣗવݴޠॲཧͷπʔϧɾσʔλ .F$BCʢܗଶૉղੳثʣ ຊޠϫʔυωοτʢγιʔϥεʣ γιʔϥεʢྨޠࣙయʣ ྨޠኮදɺຊޠޠኮେܥɺʜ ੜίʔύε ੨ۭจݿɺ8JLJQFEJBɺʜ ऍ͖ίʔύε ژେֶςΩετίʔύεɺ/"*455FYU$PSQVTɺʜ ʜ
ܗଶૉղੳث 4VEBDIJɺ+6."/ʢ+6."/ ʣɺʜ ߏจղੳث ,/1ɺ$BCP$IBɺʜ ػցֶशϥΠϒϥϦ 1Z5PSDIɺ5FOTPS'MPXɺʜ ʜ
࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ
༷ʑͳՄೳੑ Ϩετϥϯ Ϩγϐ ϝχϡʔ ޱίϛ ใݕࡧ ػց༁ จॻྨ ʜ ྉཧυϝΠϯʹ͓͚ΔԠ༻
ʹ º ྉཧυϝΠϯʹ ͓͚ΔςΩετ ࣗવݴޠॲཧͷ Ԡ༻ ʜ ใநग़ ʢͷհϖʔδʣ
Ϩγϐºใݕࡧ ͨ·Ͷ͗ πφͱͨ·Ͷ͗ͷ ෩ύελ ߜΓࠐΉ ۄೢͱੜᇙͷ ͔͋ͬͨຯो
ΧϦΧϦ̇ ΦχΦϯϦϯά ۄͶ͗ͱτϚτͷ αϥμ ΦχΦϯάϥλϯ εʔϓ ܗଶૉղੳݻ༗දݱೝࣝΛ֤Ϩγϐʹద༻͠ɺ Ϩγϐதͷ୯ޠݻ༗දݱΛೝࣝ ֤୯ޠʹ͍ͭͯಉٛޠΛల։ FH ͨ·Ͷ͗ʹۄೢʹۄͶ͗ʹΦχΦϯ ΫΤϦʢ͓Αͼɺͦͷಉٛޠʣ͕ग़ݱ͢ΔϨγϐΛ ݕࡧ݁Ռͱͯ͠ఏࣔ
Ϩγϐºจॻྨ ͨ·Ͷ͗ ओࡊ ओ৯ ۄೢͱੜᇙͷ ͔͋ͬͨຯो ෭ࡊ ो ߜΓࠐΉ
ΦχΦϯάϥλϯ εʔϓ ཛͱ౾ͱۄೢͷ ༏͍͠εʔϓ ͨ·Ͷ͗ͱਓࢀͱ Ωϟϕπͷϙτϑ ܗଶૉղੳݻ༗දݱೝࣝΛ֤Ϩγϐʹద༻͠ɺ Ϩγϐதͷ୯ޠݻ༗දݱΛೝࣝ ֤୯ޠʹ͍ͭͯಉٛޠΛల։ FH ͨ·Ͷ͗ʹۄೢʹۄͶ͗ʹΦχΦϯ ֤ΫϥεʢFH ोʣʹॴଐ͖͢ϨγϐΛਓखͰ ऩू͠ɺͦΕΒʹग़ݱ͢Δ୯ޠͷύλʔϯΛѲ Ͱऩूͨ͠ϨγϐҎ֎ͷϨγϐʹύλʔϯΛద༻͠ɺ ֤Ϋϥεʹྨ
y1 yJ y2 … canonical form of the ingredient decoder
original string of an ingredient encoder x1 x2 xI … softmax softmax softmax … Ϩγϐதͷදه ৯දதͷදه දه༳Ε ਓࢀ ʹΜ͡Μ ུه ৽ͨ· ৽ͨ·Ͷ͗ εϖϧϛε ͋΅͕Ͳ ΞϘΧυ ͦͷଞ അླḕ ͡Ό͕͍ dense amount encoder xa1 xa2 xaJ … title encoder xt1 xt2 xtI … softmax y serving λΠτϧ .ࡼ /ਓ ίίφοπΦΠϧೖΓ΄͏ΕΜΧϨʔ ࡼ ͜ΜͿɾͻ͖͡ɾ΄ͨͯ͝Μ ߹ άϥϊʔϥΫοΩʔ ຕ ύϯϓΩϯλϧτ DNλϧτܕ ৄ͘͠ɺJun Harashima, Makoto Hiramatsu, and Satoshi Sanjo. Calorie Estimation in a Real-World Recipe Service. In Proceedings of the 32nd Annual Conference on Innovative Applications of Arti fi cial Intelligence (IAAI-20), 2020. Λ͝ཡ͍ͩ͘͞ɻ Ϩγϐºػց༁ʢΧϩϦʔܭࢉʣ Ϩγϐதͷ֤ࡐྉͷදهΛ৯දதͷදهʹม ʮ.ࡼʯʮ.߹ʯɺʮ.ຕʯͳͲΛʮ/ਓʯʹม ֤ࡐྉͷΧϩϦʔΛ৯ද͔Βऔಘ͠ɺ߹ܭ ߹ܭΛ/ͰׂͬͯɺਓͷΧϩϦʔΛऔಘ ˞ผݴޠͷ༁ʢҰൠతͳػց༁ʣͰͳ͘ɺಉҰݴޠͷ༁ʢݴ͍͑ʣ
࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ
༷ʑͳՄೳੑ Χϧς จ Ϩηϓτ αϚϦ ใݕࡧ ػց༁ จॻྨ ʜ ҩྍυϝΠϯʹ͓͚ΔԠ༻
ʹ º ҩྍυϝΠϯʹ ͓͚ΔςΩετ ࣗવݴޠॲཧͷ Ԡ༻ ʜ ใநग़
จºใݕࡧ Ϩγϐºใݕࡧͱಉ͡ϩδοΫʢͷͣʣ
Χϧςºใநग़ʢ͍ΘΏΔɺϏοάσʔλͷ׆༻ʣ ܦ95&$)ʮΧϧςղੳͰਫ਼ਆ࣬ױױऀͷ༧ޙΛ༧ଌʯ IUUQTYUFDIOJLLFJDPNENBUDMGFBUVSF
ҩྍºࣗવݴޠॲཧͷ͠͞ σʔλͷॴࡏ͕Θ͔Βͳ͍ ɾ·ͩੵ͞Εͯͳ͍ʁ ɾ͞Εͯͯɺར༻ͮ͠Β͍ʁ ɾσʔλ͕ͳ͍ͱɺࣗવݴޠॲཧແྗ ɹɹɹɹσʔλͷੵɾڞ༗͕ෆՄܽ σʔλͷ༰͕Θ͔Βͳ͍ ɾΠϯεϦϯʁϔϞάϩϏϯʁܕʁܕʁ ɾࣗવݴޠॲཧʢFH ใநग़ʣͷ݁Ռ͕ղऍෆೳ
ɾղऍ͕ෆೳͩͱɺվળෆೳ ɹɹɹɹҩྍͷϓϩͱͷ࿈ܞ͕ෆՄܽ ҩྍͷૉਓ͔Βݟͨ
࣍ ࣗવݴޠॲཧͷجૅ ྉཧυϝΠϯʹ͓͚ΔԠ༻ ҩྍυϝΠϯʹ͓͚ΔԠ༻ ·ͱΊ
ࣗવݴޠॲཧͷجૅͱԠ༻ʙྉཧͱҩྍΛࡐͱͯ͠ʙ ࣗવݴޠॲཧͱʁ ɾࣗવݴޠʢʹݴ༿ʣΛίϯϐϡʔλͰѻ͏ͨΊͷٕज़ ɾ͜͜Ͱٸܹʹൃలɺଟ͘ͷاۀ͕ ςΩετσʔλ͑͋͞Εɺ༷ʑͳԠ༻͕ߟ͑ΒΕΔ ɾྉཧυϝΠϯʹ͓͚ΔςΩετºࣗવݴޠॲཧͷԠ༻ྉཧυϝΠϯʹ͓͚ΔԠ༻ ɾҩྍυϝΠϯʹ͓͚ΔςΩετºࣗવݴޠॲཧͷԠ༻ҩྍυϝΠϯʹ͓͚ΔԠ༻ ҩྍºࣗવݴޠॲཧͷ͠͞ ɾσʔλͷॴࡏ͕Θ͔Βͳ͍ˠσʔλͷੵɾڞ༗͕ෆՄܽ ɾσʔλͷ༰͕Θ͔Βͳ͍ˠҩྍͷϓϩͱͷ࿈ܞ͕ෆՄܽ