Scaling and Potential Uses for the HyperCLOVA Japanese Base Model

by Tech-Verse2022

Slide 1

Slide 1 text

No content

Slide 2

Slide 2 text

Toshinori Sato (@overlast) - Natural Language Processing - Information Retrieval Principal Software Engineer / Manager - Japanese Corpus / Evaluation / Application - mecab-ipadic-NEologd HyperCLOVA OSS: Main Contributor of NEologd ࢁ࡚ ఱ Takato Yamazaki Takato Yamazaki (@y_takaten) - Dialog Systems - HyperCLOVA Applications Software Engineer - Aizawa Lab. / University of Tokyo - “Phrase-Level Action Reinforcement Learning for Neural Dialog Response Generation”, Findings of ACL 2021 - Young Researcher Award for Excellent Research at 11th Dialogue System Symposium MS Degree in Computer Science

Slide 3

Slide 3 text

At #linedevday 2021, we presented JP 39B model

Slide 4

Slide 4 text

82B

Slide 5

Slide 5 text

Agenda Realizing Applications with Prompting Performance Evaluation of 82B JP Model Developing Applications with HyperCLOVA Deploying Foundation Models in Real World

Slide 6

Slide 6 text

Agenda Realizing Applications with Prompting Performance Evaluation of 82B JP Model Developing Applications with HyperCLOVA Deploying Foundation Models in Real World

Slide 7

Slide 7 text

Meta-learning to treat target task information as text input and predict the next word from that input Pros - Highly economical in terms of time - High accuracy can be achieved without model tuning if the appropriate information and example (few-shot) can be described Cons - Requires a considerably larger number of parameters than when using a tunable mode - Requires a linguistic background and extensive social experience Two Major Ways of Controlling Output with Foundation Model Prompting Fine-tuning Meta-learning to treat target task information as text input and predict the next word from that input Pros - Highly economical in terms of money - High accuracy can be achieved if sufficient time is available for model training and parameter exploration Cons - Need to have the right amount of properly labeled teacher data for each task - Need to create a model for each task

Slide 8

Slide 8 text

Q: But how do we use HyperCLOVA? A: Prompt Programming

Slide 9

Slide 9 text

No content

Slide 10

Slide 10 text

https://huggingface.co/spaces/stabilityai/stable-diffusion

Slide 11

Slide 11 text

; 5FDI7FSTFJTBOPOMJOFUFDIOJDBMDPOGFSFODF XIJDIXJMMCFIFMEBTBKPJOUQSPKFDUCFUXFFO-*/&BOE :BIPP+"1"/PO/PWFNCFSUIBOEUI"EJWFSTF BSSBZPGNFNCFSTGSPNUIFUXPIPTUDPNQBOJFTBOE PUIFSQBSUJDJQBUJOHPSHBOJ[BUJPOTXJMMTIBSFUIFJS DVUUJOHFEHFDIBMMFOHFTBOEBDDVNVMBUFELOPXMFEHF 1SPNQUBTBRVFSZTUSJOHPG4UBCMF%JGGVTJPO “Prompting” is becoming very common w/ Stable Diffusion

Slide 12

Slide 12 text

“Prompting” is becoming very common w/ Stable Diffusion Generate ; 5FDI7FSTFJTBOPOMJOFUFDIOJDBMDPOGFSFODF XIJDIXJMMCFIFMEBTBKPJOUQSPKFDUCFUXFFO-*/&BOE :BIPP+"1"/PO/PWFNCFSUIBOEUI"EJWFSTF BSSBZPGNFNCFSTGSPNUIFUXPIPTUDPNQBOJFTBOE PUIFSQBSUJDJQBUJOHPSHBOJ[BUJPOTXJMMTIBSFUIFJS DVUUJOHFEHFDIBMMFOHFTBOEBDDVNVMBUFELOPXMFEHF 1SPNQUBTBRVFSZTUSJOHPG4UBCMF%JGGVTJPO

Slide 13

Slide 13 text

; 5FDI7FSTFJTBOPOMJOFUFDIOJDBMDPOGFSFODF XIJDIXJMMCFIFMEBTBKPJOUQSPKFDUCFUXFFO-*/&BOE :BIPP+"1"/PO/PWFNCFSUIBOEUI"EJWFSTF BSSBZPGNFNCFSTGSPNUIFUXPIPTUDPNQBOJFTBOE PUIFSQBSUJDJQBUJOHPSHBOJ[BUJPOTXJMMTIBSFUIFJS DVUUJOHFEHFDIBMMFOHFTBOEBDDVNVMBUFELOPXMFEHF 1SPNQUBTBRVFSZTUSJOHPG4UBCMF%JGGVTJPO Hmm, this does look like a diverse array of Japanese members… “Prompting” is becoming very common w/ Stable Diffusion Generate

Slide 14

Slide 14 text

HyperCLOVA is also used with Prompt

Slide 15

Slide 15 text

"BOE#BSFDIJUDIBUUJOH HyperCLOVA is also used with Prompt 1SPNQUUPHFOFSBUFBSFTQPOTFPGBEJBMPH

Slide 16

Slide 16 text

"BOE#BSFDIJUDIBUUJOH ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF HyperCLOVA is also used with Prompt 1SPNQUUPHFOFSBUFBSFTQPOTFPGBEJBMPH

Slide 17

Slide 17 text

=== "BOE#BSFDIJUDIBUUJOH ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF HyperCLOVA is also used with Prompt 1SPNQUUPHFOFSBUFBSFTQPOTFPGUIFEJBMPH

Slide 18

Slide 18 text

=== "BOE#BSFDIJUDIBUUJOH ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF ")J XIBUEPZPVEPJOZPVSGSFFUJNF #*MJLFUPXBUDINPWJFT " 1SPNQU HyperCLOVA is also used with Prompt 1SPNQUUPHFOFSBUFBSFTQPOTFPGUIFEJBMPH

Slide 19

Slide 19 text

=== "BOE#BSFDIJUDIBUUJOH 5BTL%FTDSJQUJPO ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF %JBMPH&YBNQMFº/ ")J XIBUEPZPVEPJOZPVSGSFFUJNF #*MJLFUPXBUDINPWJFT " $VSSFOU%JBMPH HyperCLOVA is also used with Prompt *OUIJTDBTF JU`TDBMMFETIPU 1SPNQUUPHFOFSBUFBSFTQPOTFPGUIFEJBMPH

Slide 20

Slide 20 text

Slide 21

Slide 21 text

=== "BOE#BSFDIJUDIBUUJOH 5BTL%FTDSJQUJPO ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF %JBMPH&YBNQMFº/ ")J XIBUEPZPVEPJOZPVSGSFFUJNF #*MJLFUPXBUDINPWJFT " $VSSFOU%JBMPH HyperCLOVA is also used with Prompt Input the prompt and inference 1SPNQUUPHFOFSBUFBSFTQPOTFPGUIFEJBMPH

Slide 22

Slide 22 text

8IBULJOEPGNPWJFT EPZPVXBUDI === "BOE#BSFDIJUDIBUUJOH 5BTL%FTDSJQUJPO ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF %JBMPH&YBNQMFº/ ")J XIBUEPZPVEPJOZPVSGSFFUJNF #*MJLFUPXBUDINPWJFT " $VSSFOU%JBMPH Generate Input the prompt and inference HyperCLOVA is also used with Prompt 1SPNQUUPHFOFSBUFBSFTQPOTFPGUIFEJBMPH

Slide 23

Slide 23 text

8IBULJOEPGNPWJFT EPZPVXBUDI === "BOE#BSFDIJUDIBUUJOH 5BTL%FTDSJQUJPO ")J IPXBSFZPV #*`NHSFBU8IBUEPZPVEPJOZPVSGSFFUJNF ")NN *MJLFUPEPCPXMJOH #/JDF8IBU`TZPVSCFTUTDPSF ".ZCFTUTDPSFJT*VTVBMMZHFUBSPVOE #5IBU`TBXFTPNF*`MMOFWFSCFBCMFUPHFUUIBUTDPSF %JBMPH&YBNQMFº/ ")J XIBUEPZPVEPJOZPVSGSFFUJNF #*MJLFUPXBUDINPWJFT " $VSSFOU%JBMPH 8IBULJOEPGNPWJFTEPZPVXBUDI (FOFSBUFEXJUI )ZQFS$-07" HyperCLOVA is also used with Prompt Input the prompt and inference Generate 1SPNQUUPHFOFSBUFBSFTQPOTFPGUIFEJBMPH

Slide 24

Slide 24 text

That's wonderful! What do you think is the most important for your management work? OR Image Text

Slide 25

Slide 25 text

Image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utput Format Text Output Example Prompt Example Number of people who can agree on an output itself Number of people who receive an output as intended Evaluation Viewpoint Prompt is written differently depending on output modality

Slide 26

Slide 26 text

Slide 27

Slide 27 text

Slide 28

Slide 28 text

Prompt is written differently depending on output modality Image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utput Format Text Output Example Prompt Example Number of people who can agree on an output itself Number of people who receive an output as intended Evaluation Viewpoint That's wonderful! What do you think is the most important for your management work?

Slide 29

Slide 29 text

What is better output in response for user's input? === "BOE#BSFDIJUDIBUUJOH #)FZ XIBU`TVQ ")PX`TJUHPJOH "OZUIJOHOFXXJUIZPVMBUFMZ #*IFBSEUIBUUIFOFXUFOUIPVTBOEZFOCJMMXJMMIBWF 4IJCVTBXB&JJDIJPOJU CVUEPZPVLOPXBCPVUIJN " (FOFSBUFEXJUI)ZQFS$-07" 3FTQPOTFGSPNVTFS • The user's response contains a named entity “Shibusawa Eiichi” • And we can detect that “Shibusawa Eiichi” includes in the entries of Wikipedia • For simplicity, we would like to make use of the WIkipedia text after extracting “Shibusawa Eiichi” from the user's response ? "*WFCFFOSFBEJOHBMPUMBUFMZ #5IBUTHSFBU8IBUBSFZPVSFBEJOH "*WFCFFOSFBEJOHCPPLTCZ4PTFLJ/BUTVNF #)FZ *LOPXIJN5IFQFSTPOJOUIF ZFOCJMM SJHIU

Slide 30

Slide 30 text

A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: ɹɹɹɹ ? Can the training data be accurately retrieved from a model?

Slide 31

Slide 31 text

A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: I don’t really know him, who’s that? Can the training data be accurately retrieved from a model? We should assume that an accurate knowledge in the foundation model cannot be extracted. No !! Foundation Model is not a search engine !!

Slide 32

Slide 32 text

HyperCLOVA Prompt chaining can generate complex output Knowledge: Soseki Natsume was a Japanese novelist. His real name is Natsume Kinnosuke. His portrait was used on the 1,000 yen bill. Dialog: A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === Dialog: B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: "VUPHFOFSBUFE,OPXMFEHF A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: )BOENBEF,OPXMFEHF I don’t really know him, who’s that? (Something better with knowledge…) Copy Copy

Slide 33

Slide 33 text

HyperCLOVA Prompt chaining can generate complex output *IFBSEUIBUUIFOFXUFOUIPVTBOE ZFOCJMMXJMMIBWF4IJCVTBXB&JJDIJPO JU CVUEPZPVLOPXBCPVUIJN /BNFE&OUJUZ3FDPHOJUJPO

Slide 34

Slide 34 text

HyperCLOVA Prompt chaining can generate complex output *IFBSEUIBUUIFOFXUFOUIPVTBOE ZFOCJMMXJMMIBWF4IJCVTBXB&JJDIJPO JU CVUEPZPVLOPXBCPVUIJN 4IJCVTBXB&JJDIJ /BNFE&OUJUZ3FDPHOJUJPO

Slide 35

Slide 35 text

HyperCLOVA Prompt chaining can generate complex output *IFBSEUIBUUIFOFXUFOUIPVTBOE ZFOCJMMXJMMIBWF4IJCVTBXB&JJDIJPO JU CVUEPZPVLOPXBCPVUIJN /BNFE&OUJUZ3FDPHOJUJPO 4IJCVTBXB&JJDIJ %#4FBSDI FH8JLJQFEJB

Slide 36

Slide 36 text

HyperCLOVA Prompt chaining can generate complex output *IFBSEUIBUUIFOFXUFOUIPVTBOE ZFOCJMMXJMMIBWF4IJCVTBXB&JJDIJPO JU CVUEPZPVLOPXBCPVUIJN /BNFE&OUJUZ3FDPHOJUJPO 4IJCVTBXB&JJDIJ %#4FBSDI FH8JLJQFEJB Shibusawa Eiichi, 1st Viscount Shibusawa (March 16, 1840 – November 11, 1931) was a Japanese industrialist use of double-entry accounting, joint- stock corporations and modern note-issuing banks. He founded the first modern bank based on joint stock ownership in Japan. The bank was aptly named The First National Bank (Dai Ichi Kokuritsu Ginkō, now merged into Mizuho Bank) and had the power to issue its own notes. Through this bank, he founded hundreds of other joint stock corporations in Japan. Many of these companies still survive … Search Result

Slide 37

Slide 37 text

Slide 38

Slide 38 text

Slide 39

Slide 39 text

HyperCLOVA Prompt chaining can generate complex output *IFBSEUIBUUIFOFXUFOUIPVTBOE ZFOCJMMXJMMIBWF4IJCVTBXB&JJDIJPO JU CVUEPZPVLOPXBCPVUIJN Original: Natsume Sōseki (9 February 1867 – 9 December 1916), born Natsume Kin'nosuke, was a Japanese novelist. He is best known around the world for his novels Kokoro, Botchan, I Am a Cat, Kusamakura and his unfinished work Light and Darkness. He was also a scholar of British literature and writer of haiku, kanshi, and fairy tales. From 1984 until 2004, his portrait appeared on the front of the Japanese 1,000 yen note. Summary: Soseki Natsume was a Japanese novelist. His real name is Natsume Kinnosuke. His portrait was used on the 1,000 yen bill. === Original: Shibusawa Eiichi, 1st Viscount Shibusawa (March 16, 1840 – November 11, 1931) was a Japanese industrialist use of double- entry accounting, joint-stock corporations and modern note-issuing banks… Summary: Knowledge: Soseki Natsume was a Japanese novelist. His real name is Natsume Kinnosuke. His portrait was used on the 1,000 yen bill. Dialog: A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === Knowledge: Eiichi Shibusawa was born in 1840 in Kuroaraijima, Fukaya City. During his lifetime, he was involved in the establishment of approximately 500 companies and is known as the "father of Japanese capitalism”. Dialog: B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: 4VNNBSZ3FTVMU GSPN)ZQFS$-07" 4VNNBSJ[BUJPO 3FTQPOTF(FOFSBUJPO /BNFE&OUJUZ3FDPHOJUJPO 4IJCVTBXB&JJDIJ 4FBSDI3FTVMUPO 4IJCVTBXB&JJDIJ %#4FBSDI FH8JLJQFEJB

Slide 40

Slide 40 text

Slide 41

Slide 41 text

HyperCLOVA Prompt chaining can generate complex output *IFBSEUIBUUIFOFXUFOUIPVTBOE ZFOCJMMXJMMIBWF4IJCVTBXB&JJDIJPO JU CVUEPZPVLOPXBCPVUIJN Original: Natsume Sōseki (9 February 1867 – 9 December 1916), born Natsume Kin'nosuke, was a Japanese novelist. He is best known around the world for his novels Kokoro, Botchan, I Am a Cat, Kusamakura and his unfinished work Light and Darkness. He was also a scholar of British literature and writer of haiku, kanshi, and fairy tales. From 1984 until 2004, his portrait appeared on the front of the Japanese 1,000 yen note. Summary: Soseki Natsume was a Japanese novelist. His real name is Natsume Kinnosuke. His portrait was used on the 1,000 yen bill. === Original: Shibusawa Eiichi, 1st Viscount Shibusawa (March 16, 1840 – November 11, 1931) was a Japanese industrialist use of double- entry accounting, joint-stock corporations and modern note-issuing banks… Summary: Knowledge: Soseki Natsume was a Japanese novelist. His real name is Natsume Kinnosuke. His portrait was used on the 1,000 yen bill. Dialog: A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === Knowledge: Eiichi Shibusawa was born in 1840 in Kuroaraijima, Fukaya City. During his lifetime, he was involved in the establishment of approximately 500 companies and is known as the "father of Japanese capitalism”. Dialog: B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: 4VNNBSJ[BUJPO 3FTQPOTF(FOFSBUJPO /BNFE&OUJUZ3FDPHOJUJPO 4IJCVTBXB&JJDIJ 4VNNBSZ3FTVMU GSPN)ZQFS$-07" 4FBSDI3FTVMUPO 4IJCVTBXB&JJDIJ )BOENBEF4VNNBSZ %#4FBSDI FH8JLJQFEJB

Slide 42

Slide 42 text

HyperCLOVA Prompt chaining can generate complex output Natsume Sōseki (9 February 1867 – 9 r 1916), born Natsume Kin'nosuke, was se novelist. He is best known around the his novels Kokoro, Botchan, I Am a makura and his unfinished work Light ness. He was also a scholar of British and writer of haiku, kanshi, and fairy m 1984 until 2004, his portrait appeared nt of the Japanese 1,000 yen note. y: Soseki Natsume was a Japanese His real name is Natsume Kinnosuke. it was used on the 1,000 yen bill. Shibusawa Eiichi, 1st Viscount a (March 16, 1840 – November 11, s a Japanese industrialist use of double- ounting, joint-stock corporations and ote-issuing banks… y: Knowledge: Soseki Natsume was a Japanese novelist. His real name is Natsume Kinnosuke. His portrait was used on the 1,000 yen bill. Dialog: A: I've been reading a lot lately. B: That's great! What are you reading? A: I've been reading books by Soseki Natsume. B: Hey, I know him! The person in the 1,000 yen bill, right? === Knowledge: Eiichi Shibusawa was born in 1840 in Kuroaraijima, Fukaya City. During his lifetime, he was involved in the establishment of approximately 500 companies and is known as the "father of Japanese capitalism”. Dialog: B: Hey, what’s up! A: How’s it going? Anything new with you lately? B: I heard that the new ten-thousand yen bill will have Shibusawa Eiichi on it, but do you know about him? A: 4VNNBSJ[BUJPO 3FTQPOTF(FOFSBUJPO Yeah, I know him! He’s the one who laid the foundation for Japanese capitalism. He definitely deserves to be on the bill! 4VNNBSZ3FTVMU GSPN)ZQFS$-07" 4FBSDI3FTVMUPO 4IJCVTBXB&JJDIJ )BOENBEF4VNNBSZ

Slide 43

Slide 43 text

Agenda Realizing Applications with Prompting Performance Evaluation of 82B JP Model Developing Applications with HyperCLOVA Deploying Foundation Models in Real World

Slide 44

Slide 44 text

82B For 82B JP Model, we used… • Samples: 10B documents • Data size: 1.8TB • Token Size: 530B We do NOT use any data of our LINE Messenger Service • All messages on LINE • All posts on OpenChat Policy for Data Collection LINE LM Corpus Tokenizer Byte-level BPE tokenizer Library Megatron-LM Infrastructure NVIDIA Superpod 128 clustered DGX servers 1,024 A100 GPUs = Architecture Transformer Encoder-Decoder

Slide 45

Slide 45 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ HyperCLOVA w/ prompting VS BERT-large TASK: RCQA* (answerable ones only) • Removed unanswerable questions from dataset of the normal RCQA task

Slide 46

Slide 46 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ HyperCLOVA w/ prompting VS BERT-large - Questionɿதࠃ།Ұͷঁఇɺଇఱ෢޳͕ݐͯͨԦேͷ໊લ͸ԿͰ͠ΐ͏? - Contextɿ౰࣌ɺ౜Ԧே͸෢ଇఱʢଇఱ෢޳ʣʹΑΔცୣͰपԦேʹ୅Θ ͍ͬͯͨ͜ͱΛ೔ຊଆ͕೺Ѳ͍ͯ͠ͳ͔ͬͨͨΊɺ҄ాਅਓΒ͸ݱ஍Ͱ एׯͷࠞཚΛੜͨ͡ɻ·ͨɺ൴Β͕౎ɾ௕҆Ͱݟ࣮ͨࡍͷ౎৓΍཯ྩ੍ ͷӡ༻࣮ଶ͸ɺ೔ຊࠃ಺Ͱͷ૝૾ͱ͸ࣅͯඇͳΔ΋ͷͰ͋ͬͨɻͨͱ͑ ͹౻ݪژͰ͸େۃ఼ΛؚΉٶʢ౻ݪٶʣΛ౎৓ͷதԝʹ഑ஔ͍ͯͨ͠ ͕ɺ௕҆৓Λ͸͡Ίͱ͢Δதࠃͷ౎৓Ͱ͸ଠۃٶΛؚΉߖ৓͸ɺ౎৓ͷ ๺୺தԝʹ͋Δͷ͕௨ྫͰ͋ͬͨɻ཯ྩͷӡ༻ܗଶ΋೔ຊͱ͸ҟͳΓɺ ཯ྩͷෆඋΛߦ͏֨ࣜͳͲ΋੍ఆ͞Ε͍ͯͨɻେ͖ͳিܸΛड͚ͯؼࠃ ͨ҄͠ాਅਓΒ͸ɺ͜ΕΒͷ೔தͷ౎৓΍཯ྩ੍ͷࠩҟΛใࠂ͠ɺͷͪ ͷվֵʹੜ͔͞Ε͍ͯ͘ɻ Example of an entry of RCQA possible only tasks TASK: RCQA* (answerable ones only) • Removed unanswerable questions from dataset of the normal RCQA task

Slide 47

Slide 47 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ HyperCLOVA w/ prompting VS BERT-large - Questionɿதࠃ།Ұͷঁఇɺଇఱ෢޳͕ݐͯͨԦேͷ໊લ͸ԿͰ͠ΐ͏? - Contextɿ౰࣌ɺ౜Ԧே͸෢ଇఱʢଇఱ෢޳ʣʹΑΔცୣͰपԦேʹ୅Θ ͍ͬͯͨ͜ͱΛ೔ຊଆ͕೺Ѳ͍ͯ͠ͳ͔ͬͨͨΊɺ҄ాਅਓΒ͸ݱ஍Ͱ एׯͷࠞཚΛੜͨ͡ɻ·ͨɺ൴Β͕౎ɾ௕҆Ͱݟ࣮ͨࡍͷ౎৓΍཯ྩ੍ ͷӡ༻࣮ଶ͸ɺ೔ຊࠃ಺Ͱͷ૝૾ͱ͸ࣅͯඇͳΔ΋ͷͰ͋ͬͨɻͨͱ͑ ͹౻ݪژͰ͸େۃ఼ΛؚΉٶʢ౻ݪٶʣΛ౎৓ͷதԝʹ഑ஔ͍ͯͨ͠ ͕ɺ௕҆৓Λ͸͡Ίͱ͢Δதࠃͷ౎৓Ͱ͸ଠۃٶΛؚΉߖ৓͸ɺ౎৓ͷ ๺୺தԝʹ͋Δͷ͕௨ྫͰ͋ͬͨɻ཯ྩͷӡ༻ܗଶ΋೔ຊͱ͸ҟͳΓɺ ཯ྩͷෆඋΛߦ͏֨ࣜͳͲ΋੍ఆ͞Ε͍ͯͨɻେ͖ͳিܸΛड͚ͯؼࠃ ͨ҄͠ాਅਓΒ͸ɺ͜ΕΒͷ೔தͷ౎৓΍཯ྩ੍ͷࠩҟΛใࠂ͠ɺͷͪ ͷվֵʹੜ͔͞Ε͍ͯ͘ɻ - Answerɿप Example of an entry of RCQA possible only tasks TASK: RCQA* (answerable ones only) • Removed unanswerable questions from dataset of the normal RCQA task

Slide 48

Slide 48 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ HyperCLOVA w/ prompting VS BERT-large - Questionɿதࠃ།Ұͷঁఇɺଇఱ෢޳͕ݐͯͨԦேͷ໊લ͸ԿͰ͠ΐ͏? - Contextɿ౰࣌ɺ౜Ԧே͸෢ଇఱʢଇఱ෢޳ʣʹΑΔცୣͰपԦேʹ୅Θ ͍ͬͯͨ͜ͱΛ೔ຊଆ͕೺Ѳ͍ͯ͠ͳ͔ͬͨͨΊɺ҄ాਅਓΒ͸ݱ஍Ͱ एׯͷࠞཚΛੜͨ͡ɻ·ͨɺ൴Β͕౎ɾ௕҆Ͱݟ࣮ͨࡍͷ౎৓΍཯ྩ੍ ͷӡ༻࣮ଶ͸ɺ೔ຊࠃ಺Ͱͷ૝૾ͱ͸ࣅͯඇͳΔ΋ͷͰ͋ͬͨɻͨͱ͑ ͹౻ݪژͰ͸େۃ఼ΛؚΉٶʢ౻ݪٶʣΛ౎৓ͷதԝʹ഑ஔ͍ͯͨ͠ ͕ɺ௕҆৓Λ͸͡Ίͱ͢Δதࠃͷ౎৓Ͱ͸ଠۃٶΛؚΉߖ৓͸ɺ౎৓ͷ ๺୺தԝʹ͋Δͷ͕௨ྫͰ͋ͬͨɻ཯ྩͷӡ༻ܗଶ΋೔ຊͱ͸ҟͳΓɺ ཯ྩͷෆඋΛߦ͏֨ࣜͳͲ΋੍ఆ͞Ε͍ͯͨɻେ͖ͳিܸΛड͚ͯؼࠃ ͨ҄͠ాਅਓΒ͸ɺ͜ΕΒͷ೔தͷ౎৓΍཯ྩ੍ͷࠩҟΛใࠂ͠ɺͷͪ ͷվֵʹੜ͔͞Ε͍ͯ͘ɻ - Answerɿप Example of an entry of RCQA possible only tasks Create Few-shot with a context, a question text and an answer • If the correct answer is contained and easily extracted from the inference result, we judged it is correct TASK: RCQA* (answerable ones only) • Removed unanswerable questions from dataset of the normal RCQA task

Slide 49

Slide 49 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ HyperCLOVA w/ prompting VS BERT-large - Questionɿதࠃ།Ұͷঁఇɺଇఱ෢޳͕ݐͯͨԦேͷ໊લ͸ԿͰ͠ΐ͏? - Contextɿ౰࣌ɺ౜Ԧே͸෢ଇఱʢଇఱ෢޳ʣʹΑΔცୣͰपԦேʹ୅Θ ͍ͬͯͨ͜ͱΛ೔ຊଆ͕೺Ѳ͍ͯ͠ͳ͔ͬͨͨΊɺ҄ాਅਓΒ͸ݱ஍Ͱ एׯͷࠞཚΛੜͨ͡ɻ·ͨɺ൴Β͕౎ɾ௕҆Ͱݟ࣮ͨࡍͷ౎৓΍཯ྩ੍ ͷӡ༻࣮ଶ͸ɺ೔ຊࠃ಺Ͱͷ૝૾ͱ͸ࣅͯඇͳΔ΋ͷͰ͋ͬͨɻͨͱ͑ ͹౻ݪژͰ͸େۃ఼ΛؚΉٶʢ౻ݪٶʣΛ౎৓ͷதԝʹ഑ஔ͍ͯͨ͠ ͕ɺ௕҆৓Λ͸͡Ίͱ͢Δதࠃͷ౎৓Ͱ͸ଠۃٶΛؚΉߖ৓͸ɺ౎৓ͷ ๺୺தԝʹ͋Δͷ͕௨ྫͰ͋ͬͨɻ཯ྩͷӡ༻ܗଶ΋೔ຊͱ͸ҟͳΓɺ ཯ྩͷෆඋΛߦ͏֨ࣜͳͲ΋੍ఆ͞Ε͍ͯͨɻେ͖ͳিܸΛड͚ͯؼࠃ ͨ҄͠ాਅਓΒ͸ɺ͜ΕΒͷ೔தͷ౎৓΍཯ྩ੍ͷࠩҟΛใࠂ͠ɺͷͪ ͷվֵʹੜ͔͞Ε͍ͯ͘ɻ - Answerɿप Example of an entry of RCQA possible only tasks Few-shots were created randomly by extracting a context from the RCQA dev-set for each inference Create Few-shot with a context, a question text and an answer • If the correct answer is contained and easily extracted from the inference result, we judged it is correct TASK: RCQA* (answerable ones only) • Removed unanswerable questions from dataset of the normal RCQA task

Slide 50

Slide 50 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ # # # # )ZQFS$-07"+1 #&35MBSHF . HyperCLOVA w/ prompting VS BERT-large Acc. Parameters of HyperCLOVA Pre-training BERT-large by using subset of LINE LM corpus HyperCLOVA 6.9B w/ prompting is far below Fine-tuned BERT

Slide 51

Slide 51 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ # # # # )ZQFS$-07"+1 #&35MBSHF . HyperCLOVA w/ prompting VS BERT-large Acc. Parameters of HyperCLOVA HyperCLOVA 13B JP cannot exceed Fine-tuned BERT-large by only prompting

Slide 52

Slide 52 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ # # # # )ZQFS$-07"+1 #&35MBSHF . HyperCLOVA w/ prompting VS BERT-large Acc. Parameters of HyperCLOVA HyperCLOVA 39B JP is near BERT-large with fine tuning and parameter tuning by prompting

Slide 53

Slide 53 text

Automated evaluation on RCQA* possible only tasks * ղ౴Մೳੑ෇͖ಡղσʔληοτ: http://www.cl.ecei.tohoku.ac.jp/rcqa/ # # # # )ZQFS$-07"+1 #&35MBSHF . HyperCLOVA w/ prompting VS BERT-large Acc. Parameters of HyperCLOVA HyperCLOVA JP 82B with Prompting over BERT-large with full Fine-tuning and Parameter-tuning

Slide 54

Slide 54 text

Prompt Chaining is Effective for Most of the Applications Writing Business Emails Writing Business Documents Writing Novels and more…

Slide 55

Slide 55 text

Agenda Realizing Applications with Prompting Performance Evaluation of 82B JP Model Developing Applications with HyperCLOVA Deploying Foundation Models in Real World

Slide 56

Slide 56 text

Dialog System

Slide 57

Slide 57 text

Slide 58

Slide 58 text

Slide 59

Slide 59 text

• Many deployed models are rule-based • Looks for keywords to determine next response • Pros: Fully controllable • Cons: Lacks flexibility Rule-Based Dialog System )J IPX`TJUHPJOH"SFZPVSFBEZGPS5FDI7FSTF )PX`TJUHPJOH 5FDI7FSTFJTBNB[JOH *`NHSFBU Sorry, I didn’t understand the question. BOT Hey, do you like Brown? USER

Slide 60

Slide 60 text

• Many deployed models are rule-based • Looks for keywords to determine next response • Pros: Fully controllable • Cons: Lacks flexibility Rule-Based Dialog System )J IPX`TJUHPJOH"SFZPVSFBEZGPS5FDI7FSTF )PX`TJUHPJOH 5FDI7FSTFJTBNB[JOH *`NHSFBU Sorry, I didn’t understand the question. BOT Hey, do you like Brown? USER Guess what, I just got a LINE GIFT from my friend! USER

Slide 61

Slide 61 text

Slide 62

Slide 62 text

• Many deployed models are rule-based • Looks for keywords to determine next response • Pros: Fully controllable • Cons: Lacks flexibility Rule-Based Dialog System )J IPX`TJUHPJOH"SFZPVSFBEZGPS5FDI7FSTF )PX`TJUHPJOH 5FDI7FSTFJTBNB[JOH *`NHSFBU Sampl Sampl Sampl Sampl )J IPX`TJUHPJOH"SFZPVSFBEZGPS5FDI7FSTF *`NHSFBU :FT *`NTPSFBEZGPSUIFBXFTPNFQSFTFOUBUJPOT • Trending in dialog research field • Uses deep learning technique to generate response • Pros: very flexible and responses are interesting • Cons: Lacks controllability Generation-Based Dialog System

Slide 63

Slide 63 text

No content

Slide 64

Slide 64 text

Is a single prompt enough to build an amazing dialog system? %JBMPH 1SPNQU 6TFS 6UUFSBODF 4ZTUFN 3FTQPOTF

Slide 65

Slide 65 text

Probably No… A single few-shot prompt was just not enough

Slide 66

Slide 66 text

Probably No… &OEDPOWFSTBUJPO WFSZFBSMZ A single few-shot prompt was just not enough

Slide 67

Slide 67 text

Probably No… &OEDPOWFSTBUJPO WFSZFBSMZ #PSJOH SFTQPOTFT A single few-shot prompt was just not enough

Slide 68

Slide 68 text

4ZTUFN 0VUQVU 6TFS *OQVU HyperCLOVA-Based Dialog System Architecture

Slide 69

Slide 69 text

4ZTUFN 0VUQVU 6TFS *OQVU HyperCLOVA-Based Dialog System Architecture 1SFQSPDFTT 8IBUDBOXFEPIFSF 1SPNQU1SPHSBNNJOHBOE 1BSBNFUFS$POUSPMMJOH 4VCUBTLTPMWJOHXJUI )ZQFS$-07"PS/-15FDI *OTUSVDUUIF EJSFDUJPOPGPVUQVU w 4FBSDIJOHLOPXMFEHFT w %FDJEJOHEJBMPHBDU w FUD

Slide 70

Slide 70 text

Slide 71

Slide 71 text

4ZTUFN 0VUQVU 3FHFOFSBUF 6TFS *OQVU HyperCLOVA-Based Dialog System Architecture 1SFQSPDFTT 8IBUDBOXFEPIFSF 1SPNQU1SPHSBNNJOHBOE 1BSBNFUFS$POUSPMMJOH 4VCUBTLTPMWJOHXJUI )ZQFS$-07"PS/-15FDI *OTUSVDUUIF EJSFDUJPOPGPVUQVU w 4FBSDIJOHLOPXMFEHFT w %FDJEJOHEJBMPHBDU w FUD 1PTUQSPDFTT 8IBUDBOXFEPIFSF 'JMUFSJOHX/-15FDI "EEJOHBOE&EJUJOHX )ZQFS$-07"PS/-15FDI w $IFDLJOH)BMMVDJOBUJPO w $IFDLJOH&UIJDT w FUD %FUFDUPSDPSSFDU JNQSPQFSPVUQVUT (FOFSBUF $PSF0VUQVU

Slide 72

Slide 72 text

HyperCLOVA-Based Dialog System Architecture %PFTUIF LOPXMFEHFCBTF IBWFVTFGVMEBUB ,OPXMFEHF#BTFE 3FTQPOTF1SPNQU %PFTJUIBWF VTFGVMQFSTPOBEBUB 1FSTPOB$POTJTUFOU 3FTQPOTF1SPNQU (FOFSBM 3FTQPOTF1SPNQU #PSJOH3FTQPOTF %FUFDUJPO $POWFSTBUJPO&OEJOH 3FTQPOTF%FUFDUJPO .VMUJ5VSO 3FTQPOTF1SPNQU 1SPIJCJUFE8PSE %FUFDUJPO 5P3FHFOFSBUJPO 3FHFOFSBUJPO :FT :FT /P /P 1BTTFE %FUFDUFE 1BTTFE 5P3FHFOFSBUJPO %FUFDUFE %FUFDUFE 4ZTUFN 0VUQVU 6TFS *OQVU

Slide 73

Slide 73 text

Slide 74

Slide 74 text

Slide 75

Slide 75 text

Slide 76

Slide 76 text

Slide 77

Slide 77 text

With Preprocess and Postprocess Natural and engaging responses Only Few-Shot Prompt Bland and unexciting responses Comparing w/ and wo/ External Modules BOT USER BOT USER

Slide 78

Slide 78 text

With Preprocess and Postprocess Natural and engaging responses Only Few-Shot Prompt Bland and unexciting responses Comparing w/ and wo/ External Modules BOT USER BOT USER

Slide 79

Slide 79 text

Our system's achievements in dialog competitions Dialog System Live Competition 4 (2021) Dialogue system researchers gather and do live-evaluations of submitted chatbots Dialog Robot Competition 2022 A competition of controlling humanoid robot that can communicate and recommend a tourist sight inviting boss to a party Situation Track Situated Chat TU1MBDF X)ZQFS$-07"# Open Track Open-Domain Chitchat w/ random topics TU1MBDF X)ZQFS$-07"# recommending a spot Main Competition Tourist- Guide Robot TU1MBDF X)ZQFS$-07"#

Slide 80

Slide 80 text

Is that technology only for Dialog Systems?

Slide 81

Slide 81 text

HyperCLOVA Writer

Slide 82

Slide 82 text

No content

Slide 83

Slide 83 text

No content

Slide 84

Slide 84 text

Email Generation w/ HyperCLOVA Writer Email written by HyperCLOVA Writer Rough Plot by Human

Slide 85

Slide 85 text

Email Generation w/ HyperCLOVA Writer Email written by HyperCLOVA Writer Rough Plot by Human Postprocess: NER Validity Check

Slide 86

Slide 86 text

HyperCLOVA Writer is Extensible NLP APIs

Slide 87

Slide 87 text

/api/email /api/minute HyperCLOVA Writer is Extensible NLP APIs /api/novel … How can HyperCLOVA help you? Please tell us at Twitter w/ #HyperCLOVA ?????

Slide 88

Slide 88 text

What’s the goal of HyperCLOVA Writer? Human Writing Plan / Search / Write

Slide 89

Slide 89 text

What’s the goal of HyperCLOVA Writer? Human Writing Plan / Search / Write Writing w/ HyperCLOVA Plan Generate Fix Tea Time

Slide 90

Slide 90 text

What’s Next for HyperCLOVA Writer

Slide 91

Slide 91 text

Agenda Realizing Applications with Prompting Performance Evaluation of 82B JP Model Developing Applications with HyperCLOVA Deploying Foundation Models in Real World

Slide 92

Slide 92 text

82B Are foundation models perfectly ready for deployment?

Slide 93

Slide 93 text

Definitely NOT Yet 5FDI7FSTFJTBOPOMJOFUFDIOJDBMDPOGFSFODF XIJDIXJMMCFIFMEBTBKPJOUQSPKFDUCFUXFFO-*/&BOE :BIPP+"1"/PO/PWFNCFSUIBOEUI"EJWFSTF BSSBZPGNFNCFSTGSPNUIFUXPIPTUDPNQBOJFTBOE PUIFSQBSUJDJQBUJOHPSHBOJ[BUJPOTXJMMTIBSFUIFJS DVUUJOHFEHFDIBMMFOHFTBOEBDDVNVMBUFELOPXMFEHF Fairness and Ethics of Generated Text Accuracy and Reliability of Generated Text Dealing with the Cost per Inference by a Model There are still many things to think about…

Slide 94

Slide 94 text

Our First Step Towards Ethical AI 5FDI7FSTFJTBOPOMJOFUFDIOJDBMDPOGFSFODF XIJDIXJMMCFIFMEBTBKPJOUQSPKFDUCFUXFFO-*/&BOE :BIPP+"1"/PO/PWFNCFSUIBOEUI"EJWFSTF BSSBZPGNFNCFSTGSPNUIFUXPIPTUDPNQBOJFTBOE PUIFSQBSUJDJQBUJOHPSHBOJ[BUJPOTXJMMTIBSFUIFJS DVUUJOHFEHFDIBMMFOHFTBOEBDDVNVMBUFELOPXMFEHF

Slide 95

Slide 95 text

Ethics Check in Dialog System %PFTUIF LOPXMFEHFCBTF IBWFVTFGVMEBUB ,OPXMFEHF#BTFE 3FTQPOTF1SPNQU %PFTJUIBWF VTFGVMQFSTPOBEBUB 1FSTPOB$POTJTUFOU 3FTQPOTF1SPNQU (FOFSBM 3FTQPOTF1SPNQU #PSJOH3FTQPOTF %FUFDUJPO $POWFSTBUJPO&OEJOH 3FTQPOTF%FUFDUJPO .VMUJ5VSO 3FTQPOTF1SPNQU 1SPIJCJUFE8PSE %FUFDUJPO 5P3FHFOFSBUJPO 3FHFOFSBUJPO :FT :FT /P /P 1BTTFE %FUFDUFE 6TFS 6UUFSBODF 4ZTUFN 3FTQPOTF 1BTTFE 5P3FHFOFSBUJPO %FUFDUFE %FUFDUFE

Slide 96

Slide 96 text

Slide 97

Slide 97 text

Quick Fact Checking to Avoid Hallucination LINE and Yahoo! JAPAN will hold the "Tech- Verse" conference online (live streaming format) on Thursday, November 17 and Friday, November 18, 2022. Up until now, LINE and Yahoo Japan have held annual technology conferences as "LINE DEVELOPER DAY" and "Yahoo! JAPAN Tech Conference" respectively, but this year, for the first time, LINE and Yahoo Japan will hold a joint event, "Tech-Verse JAPAN Tech Conference, but this year, for the first time, LINE and Yahoo! ... Original Sentence Summarizer The overview of the sessions and speakers are available on the official website. LINE and Yahoo will jointly hold a technology conference. The theme of the conference will be "Pioneering the Future Society with Technology.” The official website, which went live today, provides an overview of the 80 sessions and speakers, with more details on the panel discussions to follow. …

Slide 98

Slide 98 text

Slide 99

Slide 99 text

Slide 100

Slide 100 text

Slide 101

Slide 101 text

Quick Fact Checking to Avoid Hallucination $POWFSUTFOUFODFTUPHSBQIT 'JOEHSBQINJTNBUDI Hallucinated

Slide 102

Slide 102 text

Human Generating Writing w/ Foundation Model Planning & Planning Writing / Searching Fixing & Fixing Generating ? Can we actually reduce labor cost by using Foundation Model?

Slide 103

Slide 103 text

Can we actually reduce labor cost by using Foundation Model? A: Depends!

Slide 104

Slide 104 text

Can we actually reduce labor cost by using Foundation Model?

Slide 105

Slide 105 text

! Good Case Plan / Write Plan Fix Gen. Can we actually reduce labor cost by using Foundation Model?

Slide 106

Slide 106 text

! Good Case Plan / Write Plan Fix Gen. " Bad Case 1 Plan / Write Plan Fix Gen. Can we actually reduce labor cost by using Foundation Model?

Slide 107

Slide 107 text

! Good Case Plan / Write Plan Fix Gen. " Bad Case 1 Plan / Write Plan Fix Gen. " Bad Case 2 Plan / Write Plan Fix Gen. × 3 Can we actually reduce labor cost by using Foundation Model?

Slide 108

Slide 108 text

Utilization of parameter-saving foundation model * JGLUE: Japanese General Language Understanding Evaluation - https://github.com/yahoojapan/JGLUE HyperCLOVA w/ LoRA tuning VS Waseda RoBERTa large {"q_id": 3016, "question": "ձࣾͷ࠷ߴ੹೚ऀΛԿͱ͍͏͔ʁ (What do you call the chief executive officer of a company?)", "choice0": "ࣾ௕ (president)", "choice1": "ڭࢣ (teacher)", "choice2": "෦௕ (manager)", "choice3": "όΠτ (part-time worker)", "choice4": "෦Լ (subordinate)", "label": 0} Example of an entry of JCommonsenseQA tasks Verification of LoRA performance of HyperCLOVA Japanese 6.7B model • Unlike Prompting, a solution is possible in 0-shot due to tuning by supervised data TASK: JCommonsenseQA • Japanese version of CommonsenseQA dataset • Five-choice QA questions to assess common sense reasoning ability This experiment was the result of a summer internship 2022 at LINE Corp. • More details will be provided at NLP2023

Slide 109

Slide 109 text

Experimental results in JCommonsenseQA HyperCLOVA w/ LoRA tuning VS Waseda RoBERTa large Human Tohoku BERT Large Waseda RoBERTa Large (s128) HyperCLOVA 6.7B w/ LoRA Accuracy 0.8 0.85 0.9 0.95 1 93.6% 90.7% 81.6% 98.8% +3.0% +14.7% -5.5% HyperCLOVA JP 6.7B with LoRA tuning exceeds JGLUE baseline accuracy

Slide 110

Slide 110 text

Challenges in Multimodal Systems

Slide 111

Slide 111 text

Tasks that had been solved by CLOVA Assistant NLP Speech Recognition Text Speech Synthesis Text

Slide 112

Slide 112 text

No content

Slide 113

Slide 113 text

This is one of our “reference” dialog system Open-domain Dialog System (e.g. For Dialog system competition 4 w/ HyperCLOVA JP 39B) Speech Recognition (LINE’s Speech-to-Text system) Speech Synthesis (LINE’s Text-to-Speech system) Text Text Voice Voice Chatting Smooth response w/ beautiful voice A dialogue system beyond the “Uncanny Valley” is imminent Wow

Slide 114

Slide 114 text

Challenge: Active handling of non-textual data in NLP NLP Speech Recognition Text Speech Synthesis Text Computer Vision Features Features Control Control Robot / Avatar

Slide 115

Slide 115 text

Challenge: Active handling of non-textual data in NLP NLP Speech Recognition Text Speech Synthesis Text Computer Vision Features Features Control Control Robot / Avatar In the future, NLP engineers need to acquire features from many modalities for NLP that other fields do not need.

Slide 116

Slide 116 text

Challenge: Active handling of non-textual data in NLP NLP Speech Recognition Text Speech Synthesis Text Computer Vision Features Features Control Control Robot / Avatar We also need to send signals to control everything that other fields are not willing to do. In the future, NLP engineers need to acquire features from many modalities for NLP that other fields do not need.

Slide 117

Slide 117 text

Foundation Model is necessary for Multi-Modal - The history of NLP is strongly linked to the development of AI-related technologies - The widespread use of the Foundation Model has made it easier to leverage the output of speech and image recognition - LINE wants to move in the direction of R&D of Multimodal NLP with our own foundation models Foundation Model DNN Traditional only ML Rule only Small LM only Text only Multi-Modal

Slide 118

Slide 118 text

Agenda Realizing applications with Prompting Performance Evaluation of 82B JP Model Developing Applications with HyperCLOVA Deploying Foundation Models in Real World When will the Japanese version of HyperCLOVA be released?

Slide 119

Slide 119 text

Slide 120

Slide 120 text

CLOVA Studio is already released for KR model

Slide 121

Slide 121 text

CLOVA Studio for JP model will come at FY2023

Slide 122

Slide 122 text

Summary - A foundation model application is already in a state to produce results that are useful enough for humans, and we are working hard to make it available to you! - However, it is necessary to prepare many subsystems to compensate for the missing functions in the foundation model - LINE is R&D these systems with the aim of providing them to you as well - LINE has successfully built a Japanese model of HyperCLOVA's 82B. We shared evaluation results and know-how for application

Slide 123

Slide 123 text

Thank you!