"医者の言葉、患者の言葉、エンジニアの言葉" / MNTSQ Ubie Vertical ai

ҩऀͷݴ༿ɺױऀͷݴ༿ɺΤϯδχΞͷݴ༿
2021/08/1
0

Ubieגࣜձࣾ @yag_ays

View Slide

2
Ԟా ༟थ

Yuki Okuda

Recruit → Sansan → Ubie

@yag_ays
ࣗݾ঺հ
https://yag-ays.github.io/

View Slide

© Ubie, Inc. 3
ҩྍػؔ޲͚ AI໰਍γεςϜ (toB) ੜ׆ऀ޲͚ ड਍૬ஊΞϓϦ (toC)
Ubie͕ఏڙ͢Δ2ͭͷϓϩμΫτ

View Slide

4
• ࣗવݴޠॲཧͷαʔϏε΁ͷԠ༻͸·ͩ·ͩϋʔυϧ͕ߴ͍Ͱ͢

• ಛʹPSF (ϓϩϒϨϜɾιϦϡʔγϣϯɾϑΟοτ) ͠͏ΔྖҬΛݟ͚ͭΔ͜ͱ͕େม

• ͦ͏ͨ͠ࢼߦࡨޡΛ͢ΔதͰߟ͑ͨ͜ͱΛ͓఻͑͠·͢

• ࣗવݴޠॲཧ΍ػցֶशΛ࢓ࣄʹ͍ͯ͠Δ/͜Ε͔Β͢Δਓʹ޲͚ͨ΢Σοτͳ࿩Ͱ͢

• ۩ମతͳٕज़Ͱ͋ͬͨΓҊ݅಺༰ͷ࿩͸͠·ͤΜ
Disclaimer
ຊࢿྉͷ಺༰͸։ൃ్தͷ಺༰Λଟؚ͘Έ·͢

View Slide

5
ҩऀɹɹͷݴ༿
ױऀ
ΤϯδχΞ

View Slide

6
ҩऀɺױऀɺΤϯδχΞ ͦΕͧΕ͕ݴ༿Λ͍࣋ͬͯΔ

View Slide

7
೤͕͋Δ
಄͕΅ʔͬͱ͢Δ
إ͕੺͍

View Slide

8
೤͕͋Δ
಄͕΅ʔͬͱ͢Δ
إ͕੺͍

View Slide

9
إ໘͕ߚை͍ͯ͠Δ
ൃ೤͍ͯ͠Δ
೤͕͋Δ
಄͕΅ʔͬͱ͢Δ
إ͕੺͍

View Slide

10
ൃ೤͍ͯ͠Δ
೤͕͋Δ
಄͕΅ʔͬͱ͢Δ
إ͕੺͍

View Slide

11
ൃ೤͍ͯ͠Δ
঱ঢ়ID:256 إ͕੺͍
පؾID:48 ෩अ
೤͕͋Δ
಄͕΅ʔͬͱ͢Δ
إ͕੺͍

View Slide

12
3ਓͷݴ༿ͷҧ͍
ҩऀͷݴ༿

• ઐ໳༻ޠͰɺ֓೦ͱݴ༿͕
໌֬ʹඥ෇͍͍ͯΔ

• ٬؍తͳ؍࡯ʹجͮ͘දݱ

• ུޠ΍ه߸ͳͲڞ௨ͷ஌ࣝ
Λ༗͍ͯ͠ΔલఏͰͷձ࿩

ױऀͷݴ༿

• ฏқͳݴ༿

• ओ؍తͳײ֮ʹجͮ͘දݱ

• දݱํ๏͕ଟ࠼ͰɺٖԻޠ
΍म০ޠΛ൐͏

ΤϯδχΞͷݴ༿

• DB΍ࣙॻʹରԠ͢Δ߲໨΍
ೖྗ಺༰ͷ໊دͤ

• ᐆດ͞Λڐ༰͠ͳ͍

• ఻ୡͷखஈΑΓɺ࣮༻ੑ΍
໢ཏੑ

View Slide

13
ͦΕͧΕͷݴ༿͸͍ͭ΍ΓऔΓ͞ΕΔ͔ʁ
ױऀ޲͚αʔϏεͷఏڙ/ར༻
໰਍΍࣏ྍ࣌ͳͲͷઆ໌

ༀࡎͷઆ໌ॻͳͲ
ҩऀ޲͚αʔϏεͷఏڙ/ར༻

View Slide

14
Ͳ͏͍ͬͨഔମͰݴ༿͕࿩͞ΕΔ͔ʁ
ױऀ޲͚αʔϏεͷఏڙ/ར༻
໰਍΍࣏ྍ࣌ͳͲͷઆ໌

ༀࡎͷઆ໌ॻͳͲ
ҩऀ޲͚αʔϏεͷఏڙ/ར༻
ిࢠΧϧςɺୀӃαϚϦʔ

͓ༀखாɺ঺հঢ়
αʔϏεͷར༻ϩά

ݚڀ΍ྟচ͔Βͷ஌ݟ
ҩऀͱͷίϛϡχέʔγϣϯ

Blog, SNSͳͲͰͷൃ৴

View Slide

15
ࣗવݴޠॲཧͷ໾ׂ͸ɺͦΕͧΕͷݴ༿Λ຋༁͢Δ͜ͱ
• ͦΕͧΕͷݴ༿ͷཪଆ

• ݴ༿ͷதʹ͸ɺͦΕΛൃͨ͠ਓͷ஌ࣝ΍ཱ৔͕҉ʹؚ·Ε͍ͯΔ

• ಉ͡೔ຊޠΛ࿩͍ͯͨ͠ͱͯ͠΋ɺ୯ޠͷҙຯ΍ද֓͢೦͕ҟͳΔ͜ͱ͕͋Δ

• ൃ৴ऀͷݴ༿ͷҙຯΛଊ͑ɺड৴ऀʹ෼͔ΔܗͰ఻͑Δ

• ױऀ͔Βҩऀɺҩऀ͔Βױऀ΁ͷ఻ୡ

• ܭࢉػ͕ͦΕΛ࣮ݱ͢Δ্Ͱ͸ɺҰ౓ΤϯδχΞͷݴ༿ʹม׵͢Δ͜ͱ͕ඞཁ

• ΤϯδχΞͷݴ༿ΛͲ͏ߏங͍͔ͯ͘͠

• ैདྷ͸ࣙॻ΍ΦϯτϩδʔͰҩྍͷߏ଄Խ͕ߦΘΕ͖͕ͯͨɺͦΕ͕͢΂ͯͰ͸ͳ͍

• ਓؒͷੈքʹର͢Δݟํͱಉ͡Α͏ʹɺΞοϓσʔτ͍͔ͯ͠ͳ͚Ε͹͍͚ͳ͍

• COVID-19ͷग़ݱ΋ͦͷҰͭ

View Slide

16
͓ༀखா΍঺հঢ়͔Βͷ৘ใநग़

View Slide

17
• ͓ༀεΩϟϯɺ঺հঢ়εΩϟϯͱ͍͏αʔϏεΛఏڙ

• ױऀ͞Μ͕࣋ࢀͨ͠ࢴഔମͷ৘ใΛεΩϟϯ͠ɺɹ
OCRʹΑΔจࣈى͜͠ͱ৘ใநग़

• ໊લͳͲͷݸਓ৘ใͷϚεΩϯά

• ͓ༀखாͷจݴͷத͔Βɺॲํ͞Ε͍ͯΔༀࡎΛநग़͢Δ

• ೔ຊͰॲํ͞Ε͍ͯΔༀࡎ໊͸ط஌ (DBͱͯ͠ଘࡏ)
͓ༀखா΍঺հঢ়͔Βͷ৘ใநग़ͱ͸
ϛϠBMࡉཻ
Χϩφʔϧৣ 仏௧࣌
ேன༦ 3แ
2ৣ
ҩऀͷݴ༿ → ΤϯδχΞͷݴ༿
※ ࢲ͕ण࢘Λ৯΂ͨ൩ʹ

ң௎Ԍʹͳͬͨͱ͖ͷॲํ

View Slide

18
͓ༀεΩϟϯͷॲཧͷྲྀΕ
1. ࢴΛεΩϟϯ͢Δ

• ը૾ͱͯ͠औΓࠐΉ

2. औಘը૾ͷิਖ਼

• ର৅ྖҬͷநग़ɺ֯౓ͷิਖ਼ɺը૾ͷ௒ղ૾

3. OCRʹΑΓจࣈͱ࠲ඪ৘ใΛಘΔ

• ෳ਺୯ҐͰग़ྗ͞ΕΔ͕ɺจࣈ୯ҐͷΈΛར༻

4. ෦෼จࣈྻΛख͕͔Γʹɺༀࡎ໊Λ࠶ߏ੒͍ͯ͘͠

• OCRಡΈऔΓϛεͷิਖ਼

• ࠶ߏ੒ͨ͠จࣈྻͱDBͷༀࡎ໊ͱͷྨࣅ౓Λࢉग़

• ޡݕग़๷ࢭͳͲͷޙॲཧΛ௥Ճ
ݱ৔ʹઃஔ͢ΔεΩϟφͱPCͷߏ੒

View Slide

19









εΩϟϯը૾͔ΒඞཁͳྖҬ΁

View Slide

20
x0 y0 x1 y1
਍ 20 12 26 18
ྍ 31 11 37 17
Պ 42 12 48 18
ɿ 56 15 62 21
಺ 63 12 69 18
Պ 72 11 78 17









࠷খ୯Ґͷจࣈͱͦͷ࠲ඪΛऔಘ͢Δ

View Slide

21









\ϒ ϓ^ \ϩ ޱ^ \ϯ ι^ \ 0 %^
Χϩφʔϧৣ200
Χϩφʔϧৣ200
Χϩφʔϧৣ300
ίΧʔϧৣ300
Χϩφʔϧࡉཻ20%
100%
90%
80%
60%
Query Documents
ༀࡎ໊ͷྨࣅ౓ʹΑΓ൑ఆ͢Δ
Α͋͘ΔಡΈऔΓϛε

View Slide

22
ϦΞϧσʔλͳΒͰ͸ͷ೉͠͞ͷྫ
1ͭͷༀࡎʹ2໊ͭલ͕هࡌ͞ΕΔ
Χϩφʔϧৣ
ҰൠɿΞητΞϛϊϑΣϯৣNH
• ઌൃༀ/ޙൃༀͷ۠ผ

• δΣωϦοΫҩༀ඼ͷ໊લͷԣʹɺݩͱ
ͳͬͨༀࡎ໊͕ซه͞ΕΔ͜ͱ͕͋Δ

• ॲํ͞Εͨༀࡎ͸1͕ͭͩɺԿ΋ߟ͑ͣʹ
நग़͢Δͱ2ͭදࣔ͞ΕΔ

• ܩઢͷͳ͍ςʔϒϧߏ଄

• ಥવͷվߦ

• OCRͱͯ͠͸ߏ଄΍จ຺Λߟྀͨ͠ಡΈ
औΓΛߦ͑ͳ͍ͷͰɺநग़จࣈྻ͕ҙਤ
ͨ͠ॱ൪ʹฒ͹ͣɺ෼அ͞ΕΔ
҉໧ͷߏ଄Խ
ϩΩιϓϩϑΣϯφճৣ೔෼
τϦ΢ϜৣNHʮ
αϫΠʯ

View Slide

23
ΞΧσϛοΫํ໘ͷಈ͖΋͋Γɺ੝Γ্͕Γͭͭ͋Δ
• Document Intelligenceͱ͍͏෼໺

• Ϗδωεจॻͷߏ଄Խͱ಺༰ཧղΛ໨తͱͨ͠ݚڀ෼໺

• NeurIPS 2019, KDD 2021Ͱworkshop͕։࠵

• ۩ମతͳλεΫ

• υΩϡϝϯτ΍ϨγʔτͷOCRɺࡱӨը૾ͷิਖ਼

• υΩϡϝϯτͷΧςΰϦࣝผ

• ςʔϒϧ΍ϑΥʔϜೝࣝͱɺ஋ͷநग़

• งғؾ

• จࣈೝࣝ΍OCRͳͲ͸ɺݩʑ͸ICDAR (International Conference on Document Analysis and
Recognition)Ͱݚڀ͞Ε͖ͯͨ

• DI͸΋͏গ͠Ϗδωε&Ԡ༻دΓ?

View Slide

24
ױऀදݱ͔Βͷ঱ঢ়நग़

View Slide

25
ױऀදݱ͔Βͷ঱ঢ়நग़ͱ͸
ױऀͷݴ༿ → ΤϯδχΞͷݴ༿
• ঱ঢ়ͱ͸

• ױऀ͕ૌ͑Δ਎ମͷҟৗ

• e.g. ಄͕௧͍ɺ଍͕ۂ͕Βͳ͍ɺԼཀྵ͕ࢭ·Βͳ͍

• ໰਍ͷೖΓޱͰ͋Γॏཁͳख͕͔Γ

• ҩऀ͕਍அΛ͢Δ্Ͱͷख͕͔Γ

• AI໰਍Ubieʹ͓͚Δ঱ঢ়ೖྗ

• Ωʔϫʔυೖྗ͔Βͷݕࡧ

View Slide

26
঱ঢ়ΛͲ͏ѻ͑͹͍͍ʁ
• ͦ΋ͦ΋঱ঢ়ͬͯԿʁ

• ۩ମͰߟ͑ͯΈΔͱ……

• ಄௧ɺෲ௧ɺൃ೤ɺԼཀྵ ͳͲ͸঱ঢ়ɻೣഎɺے೑௧͸ʁ෩अɺ΍͚Ͳɺࠎં͸ʁ

• ʮ೤͕38℃͋Δʯ͸঱ঢ়͕ͩɺʮ೤͕36℃͋Δʯ͸঱ঢ়Ͱ͸ͳ͍

• ঱ঢ়Λ໊শ໨࿥తʹهड़͢Δ͜ͱ͸ (͓ͦΒ͘) ෆՄೳ

• ଟ࠼ͳදݱ΍෇Ճ৘ใ

• ٖԻޠ/म০ޠɿ಄͕ζΩζΩ௧Ήɺӈ࿬ෲͷޙΖ͋ͨΓ͕௧Ή

• จ຺΍࣌ؒ৘ใɿ๘ʹࢗ͞Ε͔ͯΒजΕ͕ࢭ·Βͳ͍ɺٸʹ಄͕௧͘ͳͬͨ

View Slide

27
ܭࢉػ͕ѻ͑ΔܗʹͲ͏ม׵͍͔ͯ͘͠ʁ
• ਖ਼ղ͸Θ͔Γ·ͤΜ

• ߏ଄Խͨ͠ޙͷ࢖͍ํ΍λεΫʹΑͬͯɺ͓ͦΒ͘࠷దͳܗ͸มΘΓ͏Δ

• ൚༻తͳ΋ͷΛ࡞Γ͍͕ͨͰ͖ͳ͍ɺҰͭͣͭ՝୊Λղܾ͍͔ͯ͘͠͠ͳ͍

• ߟ͑ΒΕΔΞϓϩʔν

• Ξϓϩʔν1: ݴޠߏ଄ʹண໨͢Δ

• ঱ঢ়ͷදݱͷ࢓ํʹ͸ύλʔϯ͕͋ΔͩΖ͏ͱԾఆ

• ঱ঢ়ͱͯ͠࢖ΘΕΔ୯ޠ΍ͦͷ܎Γड͚ߏ଄Λݟ͚ͭύλʔϯԽ͢Δ

• Ξϓϩʔν2: ݕࡧ΍ػցֶशλεΫͱΈͳ͢

• ঱ঢ়Λ໢ཏͨࣙ͠ॻ΍Φϯτϩδʔ͸͋Δͱͯ͠ɺͦ͜ʹ౰ͯʹߦ͘

• ݕࡧͱͯ͠ͷฒͼସ͑λεΫɺ·ͨ͸ଟ஋෼ྨλεΫͱͯ͠

View Slide

28
Ξϓϩʔν1: ܎Γड͚ղੳʹΑΓओ෦/ड़෦Λൈ͖ग़͢
*1 https://acro-engineer.hatenablog.com/entry/2019/12/06/120000
• ܎Γड͚ߏ଄Λݩʹʮ಄ → ௧͍ʯͷΑ͏ͳϑϨʔζΛൈ͖ग़͢

•ʮGiNZA+ElasticsearchͰ܎Γड͚ݕࡧͷୈҰาʯΛࢀߟ *1

• ଟ͘ͷ৔߹͸ (෦Ґˠҟৗ) ΍ (ҟৗˠঢ়ଶ) ͷΑ͏ͳܗͰॻ͖ද͞ΕΔ

• ݻ༗දݱநग़ʹΑΓ෦ҐͳͲͷ෇Ճ৘ใ΋औಘ͢Δ

View Slide

29
Ξϓϩʔν2: ঱ঢ়Λ໢ཏͨࣙ͠ॻΛݩʹ༧ଌλεΫͱͯ͠ղ͘
• ઐ໳Ոͷ஌ݟ΍ݚڀ݁ՌΛݩʹɺ࣮༻తͳࣙॻ΍ΦϯτϩδʔΛ૊Έཱ͍ͯͯ͘

• Ubie͕֤࣋ͭछDB

• ݚڀྖҬͩͱHuman Phenotype Ontologyͱ͍͏Ҩ఻࣬ױ޲͚ͷදݱܕҟৗͷDB͕͋Δ

• Ͳͷදݱ͕֘౰͢Δ͔Λ༧ଌ͢ΔλεΫͱͯ͠ղ͘

• ҙຯ͕͍ۙॱʹฒͼସ্͑ͯҐn݅ΛऔΔ

• ଟ஋෼ྨͩͱΫϥε਺͕๲େ
ࡢ೔͔Β಄͕௧͍
Query
ID005: ಄ʹইΛෛͬͨ
ID006: ೤͕͋Δ
ID007: ඓ͕௧͍
ID008: ు͖ؾ͕͋Δ
ID001: ಄͕௧͍
ID002: ಄͕೤͍
ID003: ڳ͕௧͍
ID004: ޙ಄෦͕௧͍
DataBase

View Slide

30
ҩྍݴޠॲཧͱυϝΠϯ஌ࣝ

View Slide

31
ҩྍݴޠॲཧʹ͍ͭͯͷϒϩάهࣄΛॻ͖·ͨ͠
• ҩྍͱ͍͏ઐ໳ྖҬͷ஌ࣝ͸ඞཁ

• ஌͍ͬͯΔ͜ͱͰ໰୊Λఆٛ͠΍͘͢ͳΔ

• ࣙॻ΍IDମܥ͸ҩऀ΋஌Βͳ͍͜ͱ͕͋Δ

• ෼໺͕ҧͬͯ΋ख๏͸ಉ͡

• ݴޠॲཧͱ͍͏࿮૊Έ͸ಉ͡

• ҩྍͱ͍͏ಛघͳσʔλࣄ৘

• σʔλͷੑ্࣭ͳ͔ͳ͔ެ։͞Εͳ͍

• σʔλΛ࡞Δίετ͕ҟৗʹߴ͍
https://yag-ays.github.io/article/mednlp/

View Slide

32
Vertical AI Startupͱͯ͠ͷ໘ന͞
• ࣾձతՁ஋ͱݸਓతڵຯͱͷඥ෇͚

• ҩྍͱ͍͏ڀۃͷࣾձతՁ஋͕͋Δ׆ಈͱɺͦΕΛ࣮ݱ͢ΔͨΊͷελʔτΞοϓͱ͍͏৔

• ػցֶश΍ࣗવݴޠॲཧͱ͍͏ࣗ਎ͷڵຯ΍εΩϧ

• ͦͷ2ͭΛඥ෇͚Δ͜ͱͰɺࣗ਎ͷՁ஋ΛߴΊ͍ͯ͘ʢେ໊ٛ෼ͱ΋ݴ͏ʣ

• λεΫ͕ϢχʔΫͰ໘ന͍

• SOTAΛ૪͏ઓ͍Ͱ͸ͳ͘ɺ໰୊ͷϢχʔΫ͞΍αʔϏεద༻ͷՁ஋Λ໰ΘΕΔ

• ϓϨΠϠʔ͕গͳ͍͜ͱ΋͋Γɺະ։ͷ஍͕޿͕͍ͬͯΔ

• ͋ΒΏΔखஈΛۦ࢖͍ͯ͘͠૯߹ڝٕ

• ͚ͩ͜͜΍Ε͹͍͍ͱ͍͏΋ͷ͕ແ͘ɺ͋ΒΏΔྖҬΛΧόʔ͠ͳ͚Ε͹͍͚ͳ͍

• ػցֶशͷલʹɺ·ͣ͸σʔλऩू͔ΒɺͦͷσʔλΛ࡞ΔλεΫઃܭ͔Β

• ؾܰʹແྗײΛຯΘ͏͜ͱ͕Ͱ͖·͢

View Slide

33
·ͱΊ

View Slide

34
·ͱΊ
• ҩྍݴޠॲཧʹ͸3ͭͷݴ༿͕͋Δ

• ҩऀͷݴ༿ɺױऀͷݴ༿ɺΤϯδχΞͷݴ༿

• ͦΕΒͷݴ༿Λޓ͍ʹ຋༁͍ͯ͘͠ͷ͕ɺҩྍݴޠॲཧͷ໾໨

• ࣄྫ1: ͓ༀखா΍঺հঢ়͔Βͷ৘ใநग़

• ࣮༻తͳ՝୊ͳ͕Βɺݚڀ΋੝Γ্͕Γͦ͏ͳྖҬͷ࿩

• ࣄྫ2: ױऀදݱ͔Βͷ঱ঢ়நग़

• ଊ͑Ͳ͜Ζͷͳ͍ױऀͷࣗ༝ͳදݱΛ͍͔ʹߏ଄Խ͢Δ͔ʁͱ͍͏λεΫઃܭͷ࿩

• ҩྍݴޠॲཧͱυϝΠϯ஌ࣝ

• ࣾձతՁ஋ͱݸਓతڵຯͷֻ͚߹Θͤ

• ্͔ΒԼ·ͰશମΛΧόʔ͢ΔVerticalͳ؀ڥ͔ͩΒͦ͜໘ന͍

View Slide

35
We Are Hiring!!!
• ઈࢍ࠾༻தʂ

• ػցֶशΤϯδχΞ

• σʔλΤϯδχΞ

• σʔλΞφϦετ

• ιϑτ΢ΣΞΤϯδχΞ, etc.

• ·ͣ͸ؾܰʹΧδϡΞϧ໘ஊ

• ࣄۀ΍૊৫ɺ։ൃମ੍ͳͲԿͰ΋ʂ

• Ubie͸ࣾһશһ͕࠾༻׆ಈΛ͍ͯ͠·͢

• ͲͷࣾһͰ΋ྑ͍ͷͰɺ͓ؾܰʹ͓੠͕͚͘
͍ͩ͞

• https://recruit.ubie.life/engineer

https://note.com/ubie/n/n454a0d04a1eb

View Slide

"医者の言葉、患者の言葉、エンジニアの言葉" / MNTSQ Ubie Vertical ai

"医者の言葉、患者の言葉、エンジニアの言葉" / MNTSQ Ubie Vertical ai

yag_ays

More Decks by yag_ays

Other Decks in Programming

Featured

Transcript