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

LINEにおける自然言語処理技術の 研究開発の現状 / Current status of R&D of Natural Language Processing technology at LINE

LINEにおける自然言語処理技術の 研究開発の現状 / Current status of R&D of Natural Language Processing technology at LINE

LINE株式会社 佐藤敏紀 (@overlast)
第13回最先端NLP勉強会でのスポンサーセッション資料です
https://sites.google.com/view/snlp-jp/home/2021

53850955f15249a1a9dc49df6113e400?s=128

LINE Developers
PRO

September 16, 2021
Tweet

Transcript

  1. 4QPOTPST1SFTFOUBUJPOT -*/&ʹ͓͚Δࣗવݴޠॲཧٕज़ͷ 
 ݚڀ։ൃͷݱঢ় $VSSFOUTUBUVTPG3%PG/BUVSBM-BOHVBHF 1SPDFTTJOHUFDIOPMPHZBU-*/& Toshinori Sato (@overlast) LINE

    / AI Company
  2. Toshinori Sato (@overlast) • Senior Software Engineer / Manager •

    Natural Language Processing • Information Retrieval • LINE CLOVA • Japanese NLU system • HyperCLOVA • Japanese Corpus / Evaluation • OSS • Main Contributor of NEologd project • mecab-ipadic-NEologd
  3. Agenda - LINEͱ͝ࢀՃͷΈͳ͞·ͱ - LINEͰͰ͖ΔNLPٕज़ͷR&D - ࠓޙͷ LINE ͷ NLP

    ؔ࿈৘ใ
  4. Agenda - LINEͱ͝ࢀՃͷΈͳ͞·ͱ - LINEͰͰ͖ΔNLPٕज़ͷR&D - ࠓޙͷ LINE ͷ NLP

    ؔ࿈৘ใ
  5. LINEͱ͝ࢀՃͷΈͳ͞·ͱ

  6. LINE

  7. ࢖͏͚ͩͰ͸ͳ͘ɺҰॹʹ։ൃͰ͖·͢

  8. LINEͱ೔ຊͷNLPݚڀࣨͱͷؔ࿈ੑ ͋Γ͕ͱ͏͍͟͝·͢ NLPڞಉݚڀ ݱࡏ3ݚڀࣨ ଞ෼໺͔Β΋ޏ༻ NLPΞϧόΠτͷޏ༻ ݱࡏ3ਓ 2022೥΋ੋඇ NLPΠϯλʔϯ ݱࡏ1ਓ

  9. LINEͱ೔ຊͷNLPݚڀࣨͱͷؔ࿈ੑ ͋Γ͕ͱ͏͍͟͝·͢ NLPڞಉݚڀ ݱࡏ3ݚڀࣨ ଞ෼໺͔Β΋ޏ༻ NLPΞϧόΠτͷޏ༻ ݱࡏ3ਓ 2022೥΋ੋඇ NLPΠϯλʔϯ ݱࡏ1ਓ

    େֶେֶӃͰͷֶۀ΍ݚڀͷܦݧΛڧ͘ൃشͯ͠௖͍͓ͯΓ·͢
  10. NLP։ൃνʔϜͷϝϯόʔͷग़਎ɾॴଐ૊৫(Ұ෦) Ԟଜݚڀࣨ(౦޻େ) چɾদຊݚڀࣨ(NAIST) ٶඌݚڀࣨ(౦େ) ૬ᖒݚڀࣨ(NII) ༗ଜݚڀࣨ(๺େ) ࠇڮݚڀࣨ(ژେ) ߥ੉ݚڀ൝(ࡕେ) খொݚڀࣨ(౎ཱେ)

  11. NLP։ൃνʔϜͷϝϯόʔͷग़਎ɾॴଐ૊৫(Ұ෦) Ԟଜݚڀࣨ(౦޻େ) چɾদຊݚڀࣨ(NAIST) ٶඌݚڀࣨ(౦େ) ૬ᖒݚڀࣨ(NII) ༗ଜݚڀࣨ(๺େ) ࠇڮݚڀࣨ(ژେ) ߥ੉ݚڀ൝(ࡕେ) খொݚڀࣨ(౎ཱେ) ;PPNͳͲͰΧδϡΞϧʹ͓࿩Ͱ͖·͢ͷͰɺzࡶஊz͸͓ؾܰʹ͝ґཔ͍ͩ͘͞

  12. LINEͷNLPؔ࿈Πϕϯτ΁ͷεϙϯαʔࢀՃ

  13. LINEͷNLPؔ࿈Πϕϯτ΁ͷεϙϯαʔࢀՃ ࠓޙ΋/-1ίϛϡχςΟΛԠԉ͠·͢

  14. Agenda - LINEͱ͝ࢀՃͷΈͳ͞·ͱ - LINEͰͰ͖ΔNLPٕज़ͷR&D - ࠓޙͷ LINE ͷ NLP

    ؔ࿈৘ใ
  15. LINEͰͰ͖ΔNLPٕज़ͷR&D

  16. About HyperCLOVA

  17. None
  18. CLOVA Desk

  19. None
  20. LINEɾNAVER ͕ӡ༻͢Δେن໛൚༻ݴޠϞσϧ + α

  21. HyperCLOVAͷσϞ: ϝϞʹجͮ͘Ӧۀ೔ใͷ࡞੒ •ҎԼΛҰ೔ͷ๚໰ઌຖʹϦετͰ •๚໰ઌ •໨త •࣌ؒ •಺༰ •੒Ռ •௥Ճ •Ұ೔ͷ׆ಈͷཁ໿ɺ՝୊ɺཌ೔ͷ༧ఆͳͲ

  22. HyperCLOVAͷσϞ: ϝϞʹجͮ͘Ӧۀ೔ใͷ࡞੒ •ҎԼΛҰ೔ͷ๚໰ઌຖʹϦετͰ •๚໰ઌ •໨త •࣌ؒ •಺༰ɹ<= ҙ֎ͱ͕͔͔࣌ؒΔ x ๚໰਺

    •੒Ռ •௥Ճ •Ұ೔ͷ׆ಈͷཁ໿ɺ՝୊ɺཌ೔ͷ༧ఆͳͲ
  23. HyperCLOVAͷσϞ: ϝϞʹجͮ͘Ӧۀ೔ใͷ࡞੒ •ҎԼΛҰ೔ͷ๚໰ઌຖʹϦετͰ •๚໰ઌ •໨త •࣌ؒ •಺༰ɹ<= ҙ֎ͱ͕͔͔࣌ؒΔ x ๚໰਺

    •੒Ռ •௥Ճ •Ұ೔ͷ׆ಈͷཁ໿ɺ՝୊ɺཌ೔ͷ༧ఆͳͲ LINEͷϝϞఔ౓Ͱղܾͨ͠Βخ͍͠!
  24. None
  25. ֓ཁͱfew-shotͷྫ: ఆٛจཁ໿ • ֓ཁఆٛจΛཁ໿͢Δɻ • ఆٛจ҇໺ल໌ʢ͋ΜͷͻͰ͖͋ɺ೥݄೔ʣ͸ɺ೔ຊͷΞχϝʔλʔɺөը؂ಜɺ ࣮ۀՈɻΧϥʔ୅දऔక໾ࣾ௕ɻגࣜձࣾϓϩδΣΫτελδΦ2૑࡞؅ཧ౷ׅɻגࣜձࣾͰ΄͗ ΌΒΓʔऔక໾ɻ/10๏ਓΞχϝಛࡱΞʔΧΠϒػߏཧࣄ௕ɻࢁޱݝӉ෦ࢢग़਎ɻࢁޱݝཱӉ෦ ߴ౳ֶߍଔۀɻେࡕܳज़େֶܳज़ֶ෦ө૾ܭըֶՊʢݱɾө૾ֶՊʣআ੶ɻ݂ӷܕ͸"ܕɻ࠺͸ອ ըՈͷ҆໺ϞϤίɻ

    • ཁ໿҇໺ल໌͸೔ຊͷΦλΫͰ͢ɻ • ఆٛจࢳҪॏཬʢ͍ͱ͍͛͠͞ͱɺ೥ʢত࿨೥ʣ݄೔ʣ͸ɺ೔ຊͷίϐʔϥΠ λʔɺΤοηΠετɺλϨϯτɺ࡞ࢺՈɻגࣜձࣾ΄΅೔୅දऔక໾ࣾ௕<>ɻגࣜձࣾΤΠϓ୅ දऔక໾<>ɻϑΟʔϧζגࣜձࣾࣾ֎औక໾<>ɻ࠺͸ঁ༏ͷṤޱՄೆࢠɻѪݘ͸δϟοΫɾϥο ηϧɾςϦΞͷϒΠίɺ೔ຊϞϊϙϦʔڠձձ௕ɻ݂ӷܕ͸"ܕ<>ɻ਎௕DNɻ 
 • ཁ໿
  26. ֓ཁͱfew-shotͷྫ: ఆٛจཁ໿ 888 • ֓ཁఆٛจΛཁ໿͢Δɻ • ఆٛจ҇໺ल໌ʢ͋ΜͷͻͰ͖͋ɺ೥݄೔ʣ͸ɺ೔ຊͷΞχϝʔλʔɺөը؂ಜɺ ࣮ۀՈɻΧϥʔ୅දऔక໾ࣾ௕ɻגࣜձࣾϓϩδΣΫτελδΦ2૑࡞؅ཧ౷ׅɻגࣜձࣾͰ΄͗ ΌΒΓʔऔక໾ɻ/10๏ਓΞχϝಛࡱΞʔΧΠϒػߏཧࣄ௕ɻࢁޱݝӉ෦ࢢग़਎ɻࢁޱݝཱӉ෦ ߴ౳ֶߍଔۀɻେࡕܳज़େֶܳज़ֶ෦ө૾ܭըֶՊʢݱɾө૾ֶՊʣআ੶ɻ݂ӷܕ͸"ܕɻ࠺͸ອ

    ըՈͷ҆໺ϞϤίɻ • ཁ໿҇໺ल໌͸೔ຊͷΦλΫͰ͢ɻ • ఆٛจࢳҪॏཬʢ͍ͱ͍͛͠͞ͱɺ೥ʢত࿨೥ʣ݄೔ʣ͸ɺ೔ຊͷίϐʔϥΠ λʔɺΤοηΠετɺλϨϯτɺ࡞ࢺՈɻגࣜձࣾ΄΅೔୅දऔక໾ࣾ௕<>ɻגࣜձࣾΤΠϓ୅ දऔక໾<>ɻϑΟʔϧζגࣜձࣾࣾ֎औక໾<>ɻ࠺͸ঁ༏ͷṤޱՄೆࢠɻѪݘ͸δϟοΫɾϥο ηϧɾςϦΞͷϒΠίɺ೔ຊϞϊϙϦʔڠձձ௕ɻ݂ӷܕ͸"ܕ<>ɻ਎௕DNɻ 
 • ཁ໿ ֓ཁ GFXTIPU@ ೖྗ
  27. ֓ཁͱfew-shotͷྫ: ఆٛจཁ໿ • ֓ཁఆٛจΛཁ໿͢Δɻ • ఆٛจ҇໺ल໌ʢ͋ΜͷͻͰ͖͋ɺ೥݄೔ʣ͸ɺ೔ຊͷΞχϝʔλʔɺөը؂ಜɺ ࣮ۀՈɻΧϥʔ୅දऔక໾ࣾ௕ɻגࣜձࣾϓϩδΣΫτελδΦ2૑࡞؅ཧ౷ׅɻגࣜձࣾͰ΄͗ ΌΒΓʔऔక໾ɻ/10๏ਓΞχϝಛࡱΞʔΧΠϒػߏཧࣄ௕ɻࢁޱݝӉ෦ࢢग़਎ɻࢁޱݝཱӉ෦ ߴ౳ֶߍଔۀɻେࡕܳज़େֶܳज़ֶ෦ө૾ܭըֶՊʢݱɾө૾ֶՊʣআ੶ɻ݂ӷܕ͸"ܕɻ࠺͸ອ ըՈͷ҆໺ϞϤίɻ

    • ཁ໿҇໺ल໌͸೔ຊͷΦλΫͰ͢ɻ • ఆٛจࢳҪॏཬʢ͍ͱ͍͛͠͞ͱɺ೥ʢত࿨೥ʣ݄೔ʣ͸ɺ೔ຊͷίϐʔϥΠ λʔɺΤοηΠετɺλϨϯτɺ࡞ࢺՈɻגࣜձࣾ΄΅೔୅දऔక໾ࣾ௕<>ɻגࣜձࣾΤΠϓ୅ දऔక໾<>ɻϑΟʔϧζגࣜձࣾࣾ֎औక໾<>ɻ࠺͸ঁ༏ͷṤޱՄೆࢠɻѪݘ͸δϟοΫɾϥο ηϧɾςϦΞͷϒΠίɺ೔ຊϞϊϙϦʔڠձձ௕ɻ݂ӷܕ͸"ܕ<>ɻ਎௕DNɻ 
 • ཁ໿ࢳҪॏཬ͸೔ຊͷίϐʔϥΠλʔͰ͢ɻ ग़ྗ
  28. ֓ཁͱfew-shotͷྫ: ആ۟(?)ͷੜ੒ • ղઆ͔Βആ۟Λੜ੒͠·͢ɻ • */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟Ͱ͸͋Γ ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟·͟ͱ఻Θͬͯ͘ Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱ͍ͯ͠ΔقޠͰ͢ɻ • 065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ

    • */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬ͳ۟༁͸ Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ͞ʯͱʮઊͷ੠ʯͱ͍ ͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ • 065ؓ͞΍ؠʹ͠Έೖઊͷ੠ • */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ΋͜ͷٕज़Λܗʹ͢Δ ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱ͍ͯ͠Δɻ 
 065
  29. ֓ཁͱfew-shotͷྫ: ആ۟(?)ͷੜ੒ • ղઆ͔Βആ۟Λੜ੒͠·͢ɻ • */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟Ͱ͸͋Γ ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟·͟ͱ఻Θͬͯ͘ Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱ͍ͯ͠ΔقޠͰ͢ɻ • 065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ

    • */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬ͳ۟༁͸ Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ͞ʯͱʮઊͷ੠ʯͱ͍ ͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ • 065ؓ͞΍ؠʹ͠Έೖઊͷ੠ • */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ΋͜ͷٕज़Λܗʹ͢Δ ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱ͍ͯ͠Δɻ 
 065 ֓ཁ GFXTIPU@ GFXTIPU@ ೖྗ
  30. ֓ཁͱfew-shotͷྫ: ആ۟(?)ͷੜ੒ • ղઆ͔Βആ۟Λੜ੒͠·͢ɻ • */͕֝ݹ͍஑ʹඈͼࠐΜͩ࣌ͷԻͷ༷ࢠΛӵΜͩ۟Ͱ͢ɻ͕֝஑ʹඈͼࠐΉԻΛදݱͨ͠୯७ͳ۟Ͱ͸͋Γ ·͕͢ɺपғͷ੩ऐ΍ऐΕͨݹ஑ͷ༷ࢠɺ͕֝஑ʹඈͼࠐΉੜͷ༂ಈͷΑ͏ͳ৘ܠ͕·͟·͟ͱ఻Θͬͯ͘ Δɺझͷ͋Δ۟ͱͳ͍ͬͯ·͢ɻ͜ͷ۟ͷقޠ͸ʰ֝ʱͰɺ͜Ε͸य़Λදݱ͍ͯ͠ΔقޠͰ͢ɻ • 065ݹ஑΍֝ඈ͜Ήਫͷ͓ͱ

    • */ࢁܗݝʹ͋ΔཱੴࣉʢΓͬ͠Ό͘͡ʣͰηϛ͕໐͍͍ͯΔ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ͜ͷ۟͸໌֬ͳ۟༁͸ Θ͔͍ͬͯͳ͍ͷͰɺগ͠ϛεςϦΞεͰӵΈղ͘͜ͱʹϩϚϯ͕͋Δആ۟Ͱ͢ɻʮؓ͞ʯͱʮઊͷ੠ʯͱ͍ ͏Ұݟໃ६͢Δදݱͷҙਤ͕ಛʹߟ͑ͤ͞ΒΕ·͢Ͷɻ • 065ؓ͞΍ؠʹ͠Έೖઊͷ੠ • */݄຤ͷॵ͍೔ͷޕޙͷձٞʹͱͯ΋૖େͳγεςϜ։ൃͷ࿩Λ͠Α͏ͱ͍ͯ͠Δ༷ࢠΛӵ·Εͨ۟Ͱ͢ɻ ୯७ͳ۟Ͱ͸͋Γ·͕͢ɺαϥϦʔϚϯͳΤϯδχΞ͕༷ʑͳݒ೦͕͋ΔதͰɺͦΕͰ΋͜ͷٕज़Λܗʹ͢Δ ͜ͱͰਓ޻஌ೳٕज़ͷະདྷΛ։͖͍ͨͱ͍͏ر๬Λ͜Ίͨؾ࣋ͪΛදݱ͍ͯ͠Δɻ • 065݄ͷ೤͍࠭ʹຒΊͨΔΘ͕ࠢ • ˡ͜ͷޙগ͠ฤू͢Δͱྑ͍ FH೤͖݄࠭ʹຒΊͨΔΘ͕ࠢ
  31. 39BϞσϧʹର͢Δओ؍ධՁͷ࣮ࢪ ҎԼͷ༷ͳExcelϑΝΠϧΛͭ͘ΓɺධՁ؍఺ʹैͬͯ඼࣭Λ൑ఆͯ͠ɺͦͷ݁ՌͱରԠ͢Δධ఺Λॻ͘ ߦ൪߸ ࿩ऀ ςΩετ ධՁ؍఺ ධՁ؍఺ ධՁ؍఺ ධՁ؍఺ ϝϞ

    Կ͔͋Ε͹  Ϣʔβʔ これからࡶ談を開始しま す。  γεςϜ はい。わかりました。      
  32. ධՁλΠϓA: جૅޠኮͷ֬ೝ

  33. ධՁλΠϓB-1: ৚݅෇͖ͷࣗ༝࣭໰ʹର͢ΔԠ౴ΛධՁɻτϐοΫͷભҠΛ֬ೝ

  34. ධՁλΠϓB-2: τϐοΫʹର͢Δ৺৅΁ͷରԠ

  35. ධՁλΠϓC: ࣗ༝ʹࡶஊ

  36. 39BϞσϧʹର͢Δओ؍ධՁͷ࣮ࢪ Natural response Following a topic Providing a topic or

    asking a question Achievement of goals Understanding of primary and secondary vocabulary 97.8% 98.4% 0.2% 83.5% Transition between two topics 90.9% 95.3% 2.2% 90.7% Responding to users' positive feelings about a topic 90.7% 95.1% 3.3% 89.9% Responding to users' negative feelings about a topic 87.2% 93.0% 3.5% 50.5% Regular chit-chat 92.4% 93.5% 8.6% ---
  37. HyperCLOVA(େن໛൚༻ݴޠϞσϧͷڻዼ͢΂͖఺ ܰྔͳେن໛൚༻ݴޠϞσϧͱҰઢΛը͢ΔੑೳΛ࣮ݱͰ͖Δ • ख๏: େྔͷύϥϝλ(ܭࢉػ) + େྔͷσʔλ + େྔͳܭࢉྔ •

    ඞཁͳίετ = ਺ेԯԁ + ਺ඦਓͷΤϯδχΞ + 1ϲ݄ ҰઢΛը͢ΔੑೳΛ࣮ݱͰ͖ΔՄೳੑ͕ߴ͍λεΫ • ݴޠϞσϧΛ׆༻͍ͯ͠Δ • ίʔύεʹؚ·ΕΔσʔλ͕౴͑ʹͳ͍ͬͯΔ • ༩͑ͨจ຺͕໌֬ͳλεΫͷఆٛʹͳ͍ͬͯΔ
  38. HyperCLOVAΛӡ༻͢Δଆͷ༷ʑͳ՝୊ྫ • Ϟσϧͷߏங΍ɺͦͷϞσϧΛ࢖ͬͨਪ࿦ͷ࣮ࢪɺͦΕΒࣗମ͕࠷େͷνϟϨϯδ • 8FCίʔύεͷߏங • ॏෳσʔλͷ࡟আ • ར༻͢Δ֤ΤϯτϦʹؔ͢Δઆ໌੹೚ͷ࣮ݱ •

    63-*%୯ҐͰͷ࡟আཁ੥΁ͷରԠ • "*ྙཧͷ࣮૷ • ೖग़ྗԠ༻ઌʹԠͯ͡ϑΟϧλϦϯάͱཧ༝ͷ໌ࣔ • ݸਓ৘ใͷಗ໊Խ͕ඞཁͳ৔߹ • ίʔύεσʔλιʔεʹԠͯ͡ಗ໊Խ • ग़ྗঢ়گʹԠͯ͡ಗ໊Խ • ΑΓίϯύΫτͳϞσϧͷௐ੔ͱ׆༻ • Ϟσϧߏங࣌Ҏ߱ʹൃੜͨ͠৽͍͠τϐοΫ΁ͷରԠ
  39. ݴޠࢿݯ΍جૅతͳϞσϧͷߏங

  40. ༷ʑͳݴޠࢿݯ΍Ϟσϧͷߏங΋ඞཁ • λά෇͖ίʔύε • ΩʔϫʔυɾγϊχϜͷϦετ • 8FCαΠτ͝ͱͷΫϩʔϦϯάσʔλ • $PNNPO$SBXMJOH%BUBͷ࠶Ϋϩʔϧ ߋ৽

     • ஌ࣝϕʔε • ֤छͷର࿩γφϦΦ • ࢴॻ੶ͷ0$3 • ࣗಈධՁ༻ͷධՁσʔλ • খن໛ͳ൚༻ݴޠϞσϧͷվྑGJOFUVOJOHධՁ • େن໛ͳ൚༻ݴޠϞσϧͷQSPNQUJOHQSPNQUUVOJOHQBSBNFUFSUVOJOH
  41. LINE AiCall

  42. None
  43. LINE AiCall = ࣗવͳର࿩Ԡ౴Λ࣮ݱ͢ΔԻ੠ԠରAIαʔϏε

  44. ۩ମྫ1: ࣗಈԻ੠Ԡରɾ༧໿؅ཧαʔϏε

  45. ۩ମྫ1: ࣗಈԻ੠Ԡରɾ༧໿؅ཧαʔϏε NLU + Keyword Extraction => Data

  46. ۩ମྫ2: LINE௨࿩Λ༻͍ͨόΠτݕࡧαʔϏε

  47. NLU => Type of requests ۩ମྫ2: LINE௨࿩Λ༻͍ͨόΠτݕࡧαʔϏε

  48. NLU => Type of requests Datetime String Extraction => Data

    ۩ମྫ2: LINE௨࿩Λ༻͍ͨόΠτݕࡧαʔϏε
  49. ۩ମྫ3: LINE AiCallͳͲΛ༻͍ͨϫΫνϯઁऔ༧໿γεςϜ

  50. ۩ମྫ3: LINE AiCallͳͲΛ༻͍ͨϫΫνϯઁऔ༧໿γεςϜ CLOVA OCR: ը૾ܦ༝ͷςΩετ LINE AiCallͰNLU + Keyword

    Extraction
  51. Post-OCR text correction

  52. ۩ମྫ: CLOVA OCRʢϨγʔτɾྖऩॻಛԽܕʣ

  53. Ϩγʔτͷঢ়ଶ͕ѱ͍৔߹΋͋Δ => OCRΤϥʔͷൃੜ • FϑϨϯνεςΟο • ϏώμεϓϨʔϯϤʔάϧτࢷ๱η • ʔζλϧτ •

    ΞϧϓϩΦʔπ࠭౶ෆ࢖
  54. ཧ૝తʹ͸͜ͷΑ͏ʹमਖ਼͍ͨ͠(ਓؒͷߟ͑ํ) • FϑϨϯνεςΟοΫ • ϏώμεϓϨʔϯϤʔάϧτࢷ๱θϩ • νʔζλϧτ • ΞϧϓϩΦʔπ࠭౶ෆ࢖༻ ඞཁͳཁૉ:

    ର৅υϝΠϯͷσʔλͱর߹ॲཧγεςϜ
  55. ΫΤϦͱίʔύεͱͷর߹ॲཧͷͨΊͷ؍఺

  56. ྨࣅจࣈྻݕࡧϥΠϒϥϦResemblaʹΑΔর߹ॲཧ https://github.com/tuem/resembla

  57. ྨࣅจࣈྻݕࡧϥΠϒϥϦResemblaʹΑΔর߹ॲཧ https://github.com/tuem/resembla LINE CLOVAɺLINE AiCallɺ CLOVA OCR ͳͲ޿ൣғͰ׆༻

  58. LINEͷ֤छαʔϏεͷվળ

  59. $MJDL

  60. None
  61. None
  62. None
  63. None
  64. None
  65. LINEͷ֤αʔϏε΁ͷཁૉٕज़ͷಋೖ ଟ༷ͳαʔϏεͷվળʹ௅ઓͰ͖Δ ૉ๿ͳNLPతॲཧͰߏ੒͞Εͨػೳ͸࿙Εͳ͘ର৅ 2021೥͸NLP͕ແ͍ͱ੒ཱ͠ͳ͍ػೳ΍αʔϏεʹଟ ͘ͷϢʔβʔ΍։ൃऀͷؔ৺͕޲͍͍ͯΔ ɹ- 2012೥ࠒ͸ͦ͏Ͱ΋ແ͔ͬͨ ɹɹ- શࠃେձͷεϙϯαʔϦετͰ΋໌Β͔ ଘࡏ͠ͳ͍ٕज़ͷ࣮ݱํ๏͸ࣗ෼࣍ୈ

    ɹ- ઌ୺ݚڀͱଈ࣮༻ɺ൚༻ͱಛԽɺࣗಈԽͱਓख 
 ɹ- ೔ຊޠͱଟݴޠɺςΩετͱϚϧνϞʔμϧ 
 ɹ- ͳͲɺߟྀ͢Δ΂͖؍఺΋ଟ͍ 

  66. LINEͰͰ͖ΔNLPٕज़ͷR&Dͷ·ͱΊ - LINEͷB2B: ྫɺLINE AiCall΍CLOVA OCR - LINEͷ֤αʔϏε΁ͷཁૉٕज़ͷಋೖ - LINEͷ೔ຊޠʹಛԽͨ͠େن໛൚༻ݴޠϞσϧͷߏஙɾධՁ

    - େن໛൚༻ݴޠϞσϧͷӡ༻ʹඞཁͳཁૉٕज़ͷ։ൃ - ֤छݴޠࢿݯͷ࡞੒ - ࿦จࣥච / ڞ௨λεΫ΁ͷࢀՃ / OSSͱͯ͠ެ։ɾߩݙ
  67. Agenda - LINEͱ͝ࢀՃͷΈͳ͞·ͱ - LINEͰͰ͖ΔNLPٕज़ͷR&D - ࠓޙͷ LINE ͷ NLP

    ؔ࿈৘ใ
  68. ࠓޙͷ LINE ͷ NLP ؔ࿈৘ใ

  69. None
  70. ࠓ೥ͷʙʹΦϯϥΠϯ։࠵͠·͢ɻΑΖ͚͠Ε͹͝ࢀՃ͍ͩ͘͞

  71. https://linecorp.com/ja/career/

  72. https://linecorp.com/ja/career/

  73. https://linecorp.com/ja/career/

  74. None
  75. ೥౓ͷ৽ଔ࠾༻ʹ͍ͭͯ΋্݄०Φʔϓϯ͠·͢ͷͰɺੋඇ͝Ԡื͍ͩ͘͞

  76. ·ͱΊ - LINE͸͝ࢀՃͷΈͳ͞·ͱNLP͕͍ͨ͠Ͱ͢ - LINEͰNLPͷR&DͰ͖·͢ - ࠓޙ΋ LINE ͸ NLP

    ؔ࿈৘ใͷൃ৴͠·͢
  77. See you NLP2022 !!