Slide 1

Slide 1 text

ୈ14ճ࠷ઌ୺NLPษڧձ εϙϯαʔυηογϣϯߨԋ Toshinori Sato (@overlast) LINE / AI Company 

Slide 2

Slide 2 text

ࠤ౻ හل (@overlast) ● γχΞιϑτ΢ΣΞΤϯδχΞ / Ϛωʔδϟʔ ● ࣗવݴޠॲཧ ● ৘ใݕࡧ ● LINE CLOVA ● ೔ຊޠͷࣗવݴޠཧղ(NLU)γεςϜ ● HyperCLOVA ● ೔ຊޠίʔύε / AIϑΟϧλʔ / Ԡ༻ٕज़ͷ։ൃɾධՁ ● OSS: NEologd ϓϩδΣΫτͷϝΠϯ։ൃऀ ● mecab-ipadic-NEologd 2

Slide 3

Slide 3 text

NLP։ൃνʔϜʹ͍ͭͯ

Slide 4

Slide 4 text

2012೥5݄ (1ਓ)

Slide 5

Slide 5 text

2022೥9݄ (20ਓ) 1݄ ٢઒ࠀਖ਼͞Μ 5 5݄ ਫຊஐ໵͞Μ 8݄ɹਗ਼໺ॢ͞Μ 9݄ ೳ஍޺͞Μ ࠓޙ΋Ұॹʹಇ͚ͯخ͍͠஥͕ؒՃΘΔ༧ఆ

Slide 6

Slide 6 text

No content

Slide 7

Slide 7 text

νʔϜ಺Ϣχοτ੍Λ࠾༻ ཁૉٕज़ɾϓϥοτϗʔϜ։ൃϢχοτ - ཁૉٕज़Λݚڀ։ൃ͠ɺ࣮૷ΛαʔϏεಋೖ΍ϓϥοτϗʔϜԽɾ044Խ͢Δ ج൫ϞσϦϯάϢχοτ - ͓΋ʹ೔ຊޠͷ͍ΘΏΔେن໛൚༻ݴޠϞσϧͷߏஙɾධՁɾαʔϏεԠ༻Λ࣮ࢪ ର࿩γεςϜ։ൃϢχοτ - ચ࿅͞Εͨର࿩γεςϜΛߏ੒͢ΔͨΊʹඞཁͳ༷ʑͳٕज़Λݚڀ։ൃ͢Δ Ϗδωε։ൃϢχοτ - ࣾ಺֎ͷ##ɾ#$αʔϏε΁ͷٕज़ఏҊɾಋೖͳͲΛ࣮ࢪ͢Δ

Slide 8

Slide 8 text

About HyperCLOVA

Slide 9

Slide 9 text

No content

Slide 10

Slide 10 text

೔ຊޠͷج൫Ϟσϧ + α ͳγεςϜ Automatic evaluation with 39B JP Model for a QA task

Slide 11

Slide 11 text

HyperCLOVA ͷ೔ຊޠϞσϧߏஙͷݱঢ় 1.3B → 6.7B → 13B → 39B 13B → 39B 82B ଟݴޠϞσϧ େن໛Ϟσϧ 
 ೔ຊޠ / ଟݴޠ ௒େن໛ 
 ೔ຊޠϞσϧ ೔ຊޠϞσϧ ߏங࡞ۀ͕ਐߦத  - Ϟσϧߏஙͷ೉͠͞ : ࣮Ԡ༻͕͏·͍͘͘ɺͱ෼͔Δ·ͰऴΘΒͳ͍ - ࣄલ४උ(Πϯϑϥɾσʔλ) → Ϟσϧߏங → ϞσϧධՁ → ϞσϧΛԠ༻ͨ݁͠ՌͷධՁ

Slide 12

Slide 12 text

HyperCLOVAͷσϞ: 3 ͭͷλεΫΛಉ͡ϞσϧͰ HyperCLOVAͷՄೳੑΛ֬ೝͯ͠௖ͨ͘ΊͷσϞͰ͢ ϓϨεϦϦʔε ͷੜ੒ཁ໿ ϖϧιφΛՃຯͰ͖ͨ γφϦΦແ͠ͷ 
 ର࿩γεςϜ ! ࡶͳϝϞΛૹ৴Մೳঢ়ଶʹ ओ؍ධՁ݁Ռ͕ྑ͍  ϝʔϧͷੜ੒

Slide 13

Slide 13 text

No content

Slide 14

Slide 14 text

ίʔύεͷͨΊͷσʔλऩूͷํ਑ LINEͷαʔϏεͷձ࿩σʔλΛ࢖Θͳ͍ - LINEͷ͢΂ͯͷϝοηʔδ - OpenChatͷ͢΂ͯͷ౤ߘ ݖརॲཧΛਖ਼͓͘͜͠ͳ͔ͬͯΒίʔύεʹ௥Ճ͢Δ  ൚༻ੑΛ޲্Ͱ͖ΔίϯςϯπΛඃ෴͢Δ༷ʹΫϩʔϧ͢Δ - LINEͷ֎ʹίʔύεͷαϒηοτΛఏڙ͢ΔՄೳੑΛ࢒͢ 

Slide 15

Slide 15 text

 Λࣗ༝ʹ࢖͑Δͱݚڀతʹ͸Կ͕خ͍͠ͷʜʁ

Slide 16

Slide 16 text

 ࠷ॳͷσϞ͕ग़དྷ্͕Δ·Ͱͷ 
 εϐʔυײ͕͍͢͝

Slide 17

Slide 17 text

σϞ࡞੒·Ͱͷεϐʔυײʂ  ͜ͷ໰୊)ZQFS$-07"Ͱ 
 ղ͚ͦ͏͡Όͳ͍ʜʁ ΍ͬͯΈΑ͏ʂ ͜ͷϓϩηε͕ͨͬͨ਺೔ʂ ࠷ॳͷσϞ͕Ͱ͖·ͨ͠

Slide 18

Slide 18 text

 ΩϟϥΫλʔͱͷ νϟοτΛ࣮ݱ͍ͨ͠ʂ ྫ͑͹ʜ

Slide 19

Slide 19 text

ࠓ·Ͱͷݚڀ։ൃ͸ʜ  ΩϟϥΫλʔର࿩ʢϖϧιφର࿩ʣσʔλΛूΊΔ Ϟσϧߏஙɾ܇࿅ɾύϥϝλ୳ࡧ σϞ࡞੒ αʔϏεاը ͜͜ʹ͍ͬͺ͍ ͕͔͔࣌ؒΔ ࣮ݱՄೳੑ͕Θ͔Δ·Ͱ͕͔͔࣌ؒΔ

Slide 20

Slide 20 text

)ZQFS$-07"͕࢖͑ΔͳΒʜ  σϞ࡞੒ αʔϏεاը ͜ΕͳΒ ͙͢ʹͰ͖Δʂ ϓϩϯϓτϓϩάϥϛϯά

Slide 21

Slide 21 text



Slide 22

Slide 22 text



Slide 23

Slide 23 text



Slide 24

Slide 24 text

εϐʔυ͕ૣ͍ͱԿ͕ྑ͍ͷʁ  ࣮ݱՄೳੑ͕͙͢Θ͔ΓɺϚωδϝϯτ૚΍Ӧۀ෦ॺ΁ͷઆಘྗ͕͍͢͝ ˣ͜ΜͳσϞ͕൒೥Ͱ ݸͰ͖Δʢؾ߹ೖΕΕ͹ଟ෼΋ͬͱͰ͖Δʣ ےͷྑ͍՝୊͔ɺͲͷఔ౓ͷ೉қ౓ͷ໰୊͔ɺͷ൑அ͕ૣ͘ͳΔ

Slide 25

Slide 25 text

LINEͷNLP։ൃνʔϜ͸2022೥΋ྭΈ·͢ ࣾ಺֎ͷ༷ʑͳαʔϏεʹରͯ͠ߩݙΛ͢ΔͨΊͷۀ຿͕جຊ ɹ- ϚϧνϞʔμϧԽͳͲʹΑΔNLPͰѻ͏ྖҬͷ֦େ͕՝୊ ର࿩ϩϘοτίϯϖςΟγϣϯͷຊબձʹग़৔(༧બ2Ґ௨ա) ର࿩γεςϜϥΠϒίϯϖςΟγϣϯͷ༧બʹग़৔༧ఆ ༷ʑͳ੒Ռɾ஌ݟɾOSSΛ͓ಧ͚͢ΔͨΊʹӶҙ։ൃத ɹ- ग़લؗͷݕࡧΫΤϦͷमਖ਼ͳͲ͸࠷ۙͷେ͖ͳ੒Ռͷͻͱͭ