[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] Emoberta: Speaker-aware emotion recognition in conversation with roberta
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Kim, Taewoon, and Piek Vossen. "Emoberta: Speaker-aware emotion recognition in conversation with roberta." arXiv preprint arXiv:2108.12009 (2021). introduction ๊ฐ์ • ์ธ์‹์˜ ๋ฒ”์œ„๋Š” ํ‘œ์ •, ์Œ์„ฑ,ํ…์ŠคํŠธ ๋“ฑ ๋งค์šฐ ๋„“์Œ. ๋ณธ ๋…ผ๋ฌธ์—์„œ๋Š” ๊ฐ์ • ์ธ์‹์˜ ํ•˜์œ„ ๋ถ„์•ผ์ธ ๋Œ€ํ™”์—์„œ ๊ฐ์ • ์ธ์‹( ERC)์— ์ค‘์ ์„ ๋‘ . ERC๋Š” ํ•œ์‚ฌ๋žŒ ๋˜๋Š” ๊ทธ ์ด์ƒ์˜ ์‚ฌ๋žŒ๊ณผ ๋Œ€ํ™”์— ์ฐธ์—ฌํ•˜๋Š” ํ˜„์žฌ ํ™”์ž์˜ ๊ฐ์ •์„ ์˜ˆ์ธกํ•˜๋Š” ๊ฒƒ ๊ฐ์ •์„ ์ธ์‹ํ•˜๋Š” ๊ฒƒ์€ ๊ฐ์„ฑ ์ปดํ“จํŒ… ๋ฐ ์ธ๊ฐ„-๋กœ๋ด‡ ํ†ต์‹ ๊ณผ ๊ฐ™์€ ๋ถ„์•ผ์—์„œ ์ค‘์š”. ์ธ๊ฐ„์€ ๋Œ€ํ™”(์‹œ๊ฐ, ์Œ์„ฑ ๋“ฑ)์„ ํ•˜๊ธฐ์œ„ํ•ด ์—ฌ๋Ÿฌ ๊ฐ๊ฐ ์ž…๋ ฅ์„ ์‚ฌ์šฉํ•จ ๋”ฐ๋ผ์„œ ERC ์ž‘์—…์—๋Š” ์—ฌ๋Ÿฌ ์–‘์‹(์‹œ๊ฐ, ์˜ค๋””์˜ค, ํ…์ŠคํŠธ..
๊ฐ์ •์„ ๋…น์ธ KoGPT ๊ธฐ๋ฐ˜ ์ฑ—๋ด‡ ์ œ์ž‘
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
์ฃผ์ œ : ์ค‘๊ณ ๋“ฑํ•™์ƒ์„ ์œ„ํ•œ ํ•™๊ณผ ๊ต์ˆ˜๋‹˜ ์ธํ„ฐ๋ทฐ ์ฑ—๋ด‡ ์˜์˜ : ๋Œ€ํ•™ ์ง„ํ•™ ์ „, ์œ ์šฉํ•œ ์ •๋ณด๋ฅผ ์–ป์„ ์ˆ˜ ์žˆ์Œ! ๊ฒฐ๊ณผ ๋ฏธ๋ฆฌ๋ณด๊ธฐ : ํ•™์Šต์— ํ•„์š”ํ•œ ๋ฐ์ดํ„ฐ์…‹ : ํ•™๊ณผ์˜ ๊ฐœ์ˆ˜๋Š” ์ด 320๊ฐœ๊ฐ€ ์žˆ์œผ๋ฉฐ, ํ‰๊ท  ์งˆ์˜์‘๋‹ต ๊ฐœ์ˆ˜๋Š” 11์Œ, ์ตœ๋Œ€ 19์Œ ๊นŒ์ง€ ์ด์–ด์ง‘๋‹ˆ๋‹ค. ๋ฐ์ดํ„ฐ์…‹์˜ ๋Œ€ํ‘œ ์งˆ๋ฌธ: 1.๊ต์ˆ˜๋‹˜๊ป˜์„œ ์ „๊ณต ์„ ํƒํ•œ ๋™๊ธฐ์ด ๋ฌด์—‡์ธ๊ฐ€์š” 2.์–ด๋–ค ๋ถ„์•ผ์— ๊ด€์‹ฌ ๊ฐ–๋Š” ์‚ฌ๋žŒ์ด ์ด ํ•™๊ณผ๋กœ ์˜ค๋ฉด ์ข‹์„๊นŒ์š” 3.์ด ํ•™๊ณผ์—์„œ ๊ฐ€์žฅ ์ค‘์š”ํ•œ ๊ณต๋ถ€๋Š” ์–ด๋–ค ๋‚ด์šฉ์ธ๊ฐ€์š” 4.์ด ํ•™๊ณผ์—์„œ ๊ณต๋ถ€ ์ž˜ํ• ๋ ค๋ฉด ์ค‘๊ณ ๋“ฑํ•™์ƒ๋•Œ ์–ด๋–ค ๊ต๊ณผ๋ชฉ์„ ๊ณต๋ถ€ํ•˜๋ฉด ์ข‹์€๊ฐ€์š” 5.์ด ํ•™๊ณผ์˜ ์žฅ์ ์€ ๋ญ”๊ฐ€์š” 6.์ด ํ•™๊ณผ์˜ ํ•™์ƒ๋“ค์ด ๊ฒช๋Š” ์–ด๋ ค์›€์€ ๋ญ”๊ฐ€์š” 7.ํ•™๊ณผ ์กธ์—…์ƒ์ด ๊ฐ€์žฅ ๋งŽ์ด ์ง„์ถœํ•˜๋Š” ์ง์—…๋ถ„์•ผ๋Š” ์–ด๋Š ๊ณณ์ธ๊ฐ€์š” 8.ํ•™๊ณผ์˜ ์•ž์œผ๋กœ์˜ ์ „๋ง์€ ์–ด๋–ค๊ฐ€์š” 9.์ง€๊ธˆ์€ ์—†์ง€๋งŒ ์•ž์œผ๋กœ ์ƒˆ๋กœ ์ƒ๊ธฐ๊ฒŒ..
[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] GPT Understands, Too
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Liu, Xiao, et al. "GPT understands, too." arXiv preprint arXiv:2103.10385 (2021). NLU Task์—์„œ ์•ฝํ•จ์„ ๋ณด์—ฌ์ฃผ์—ˆ๋˜ GPT๊ฐ€ BERT ๊ณ„์—ด์„ ์ด๊ธด P-Tuning์— ๋Œ€ํ•œ ๋…ผ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ด๋ฅผ ์ดํ•ดํ•˜๊ธฐ ์œ„ํ•ด์„œ ๊ทธ๋™์•ˆ์˜ ์ž์—ฐ์–ด ์ดํ•ด ์—ฐ๊ตฌ ํ๋ฆ„์— ๋Œ€ํ•ด์„œ๋„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค.
[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] FEDNLP, Federated Distillation of Natural Language Understanding with Confident Sinkhorns
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Lin, Bill Yuchen, et al. "Fednlp: Benchmarking federated learning methods for natural language processing tasks." arXiv preprint arXiv:2104.08815 (2021). Abstract ์ž์—ฐ์–ด ์ฒ˜๋ฆฌ(NLP) ์ž‘์—…์„ ์œ„ํ•œ ๊ฐœ์ธ ์ •๋ณด ๋ณดํ˜ธ, ๋ถ„์‚ฐํ˜• ํ•™์Šต ๋ฐฉ๋ฒ•์— ๋Œ€ํ•œ ์—ฐ๊ตฌ๊ฐ€ ํ•„์š” NLP ์ž‘์—…์— ๋Œ€ํ•œ FL ๋ฐฉ๋ฒ• ์—ฐ๊ตฌ์— ๋Œ€ํ•œ ๊ด€์‹ฌ์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  ๋ฌธํ—Œ์—๋Š” ์ฒด๊ณ„์ ์ธ ๋น„๊ต ๋ฐ ๋ถ„์„์ด ๋ถ€์กฑ ํ…์ŠคํŠธ ๋ถ„๋ฅ˜, ์‹œํ€€์Šค ํƒœ๊น…, ์งˆ๋ฌธ ๋‹ต๋ณ€ ๋ฐ seq2seq์˜ ๋„ค ๊ฐ€์ง€ ์ž‘์—… ๊ณต์‹์— ๋Œ€ํ•œ ์—ฐํ•ฉ ํ•™์Šต ๋ฐฉ๋ฒ•์„ ํ‰๊ฐ€ํ•˜๊ธฐ ์œ„ํ•œ ๋ฒค์น˜๋งˆํ‚น ํ”„๋ ˆ์ž„์›Œํฌ์ธ FedNLP๋ฅผ ์ œ์‹œ Motivate FL ์˜์—ญ์˜ ๋ฐœ์ „์—๋„ ๋ถˆ๊ตฌํ•˜๊ณ  NLP์— ๋Œ€ํ•œ ์—ฐ๊ตฌ์™€ ์ ์šฉ..
[๋…ผ๋ฌธ๋ฆฌ๋ทฐ] Federated split bert for heterogeneous text classification
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Federated split bert for heterogeneous text classification Lit, Zhengyang, et al. "Federated split bert for heterogeneous text classification." 2022 International Joint Conference on Neural Networks (IJCNN). IEEE, 2022. Abstract ์—ฐํ•ฉ ํ•™์Šต ์„ค์ •์—์„œ BERT๋ฅผ ์กฐ์‚ฌํ•œ ๋ช‡ ๊ฐ€์ง€ ์—ฐ๊ตฌ๊ฐ€ ์žˆ์ง€๋งŒ ํด๋ผ์ด์–ธํŠธ์— ๋Œ€ํ•œ ์ด๊ธฐ์ข…(์˜ˆ: ๋น„ IID) ๋ฐ์ดํ„ฐ๋กœ ์ธํ•œ ์„ฑ๋Šฅ ์†์‹ค ๋ฌธ์ œ๋Š” ์•„์ง ์ œ๋Œ€๋กœ ์กฐ์‚ฌ๋˜์ง€ ์•Š์Œ. ์ด ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ธฐ ์œ„ํ•ด BERT ์ธ์ฝ”๋” ๊ณ„์ธต์„ ๋กœ์ปฌ ๋ถ€๋ถ„๊ณผ ์ „์—ญ ๋ถ€๋ถ„์œผ๋กœ ๋ถ„ํ• ํ•˜์—ฌ ์ด์ข… ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ณ  ํ†ต์‹  ๋น„์šฉ์„ ์ค„์ด๋Š” ํ”„๋ ˆ์ž„์›Œํฌ์ธ..
To be uploaded Papers list
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Liu, Xiao, et al. "P-Tuning: Prompt Tuning Can Be Comparable to Fine-tuning Across Scales and Tasks." Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) . 2022. P-tuning Ver.2์— ๊ด€ํ•œ ๋…ผ๋ฌธ์ด๊ณ , ๋ชจ๋ธ scale์ด๋‚˜ NLU Task์— ๊ด€๋ จ์—†์ด ์ตœ์ ํ™”๋œ Prompt ๊ตฌ์กฐ ์ œ์‹œ Masked Language Modeling ๋ฟ๋งŒ ์•„๋‹ˆ๋ผ Classification์—์„œ ์ ์šฉ ๊ฐ€๋Šฅํ•ด์ง Deep Prompt Tuning ๊ตฌ์กฐ ์ฑ„ํƒ ๋ชจ๋“  ๋ ˆ์ด์–ด์— Continuous promp..
Recent Natural Language Paper Flows
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
์ž์—ฐ์–ด ๋…ผ๋ฌธ์˜ ํ๋ฆ„๋„ ์ด์ „๊นŒ์ง€ ์ž์—ฐ์–ด์ฒ˜๋ฆฌ๋Š” ๊ธฐ๊ณ„ํ•™์Šต์—์„œ RNN์˜ ๋“ฑ์žฅ์œผ๋กœ์ธํ•ด ์ ์ฐจ ๋”ฅ๋Ÿฌ๋‹์œผ๋กœ ๋“ค์–ด๊ฐ€๊ฒŒ ๋˜์—ˆ์Œ. RNN ์ˆœํ™˜์‹ ๊ฒฝ๋ง์—์„œ vanishing gradient ๋ฌธ์ œ๊ฐ€ ์žˆ์–ด์„œ, ๊ธฐ์–ตshell์„ ์ถ”๊ฐ€ํ•œ LSTM์ด ๋‚˜์™”์œผ๋ฉฐ, ๊ทธ๋’ค๋กœ GRU๋„ ์“ฐ๊ณ  ํ•˜๋‹ค๊ฐ€ ํŠธ๋žœ์Šคํฌ๋จธ๊ฐ€ ๋‚˜์˜จ ๋’ค๋กœ๋ถ€ํ„ฐ ์–˜๊ฐ€ ์ด์ „๊นŒ์ง€์˜ ์„ฑ๋Šฅ์„ ๋‹ค ์ด๊ฒจ๋ฒ„๋ ค์„œ ์ž์—ฐ์–ด์ฒ˜๋ฆฌ์˜ ํ๋ฆ„์€ ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ์„ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•œ ์—ฐ๊ตฌ๊ฐ€ ๋งŽ์•„์กŒ์Œ. Attention Is All You Need Vaswani, Ashish, et al. "Attention is all you need." Advances in neural information processing systems 30 (2017). ์–ดํ…์…˜ ๊ธฐ๋ฒ• ์†Œ๊ฐœ → ํŠธ๋žœ์Šคํฌ๋จธ ๋ชจ๋ธ On. NLP ๋ถ„์•ผ์—์„œ ์ „์„ค์˜ ์‹œ์ž‘ ..
[๋…ผ๋ฌธ๊ตฌํ˜„]Self-Explaining Structures Improve NLP Models
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Sun, Zijun, et al. "Self-explaining structures improve nlp models." arXiv preprint arXiv:2012.01786 (2020). ํ˜„์žฌ ์ž์—ฐ์–ด ์ถ”๋ก ๋ถ€๋ถ„์—์„œ ์ƒ์œ„ ๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•˜๊ณ  ์žˆ๋Š” ๋…ผ๋ฌธ์ด๋‹ค. ์ฒ˜์Œ Pre-print๋ผ๋Š” ๊ฒƒ์„ ํ™•์ธํ–ˆ์„ ๋•Œ ๊ผฌ๋ฆ„ํ•œ๊ฑธ ์•Œ์•„์ฐจ๋ ธ์–ด์•ผํ–ˆ๋Š”๋ฐ, ์—ฌ๋Ÿฌ ๋Œ€ํšŒ์—์„œ๋„ ๋ณธ ๋…ผ๋ฌธ์˜ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•œ ๋ฐฉ๋ฒ•๋„ ๋งŽ๊ณ , ์ ์ˆ˜๋„ ๋†’๊ฒŒ๋‚˜์™€์„œ ์ข‹์€ ๋ฐฉ๋ฒ•์ด๊ตฌ๋‚˜ ํ•˜์˜€๋‹ค. ํ•™ํšŒ์— ๋“ฑ๋ก ์•ˆ๋œ๊ฑธ ๋ฌด์‹œํ•˜๋ฉด ์•ˆ๋˜์—ˆ๋‹ค.. ๋…ผ๋ฌธ ๊ทธ๋Œ€๋กœ ์ง์ ‘ ๊ตฌํ˜„ํ•ด๋ณด๋‹ˆ Base ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋” ์•ˆ์ข‹๋‹ค. ์ด ๊ธฐ์ˆ  ์‚ฌ์šฉํ•œ๋‹ค๊ณ  ๋ช‡์ฃผ๋‚ ๋ฆฐ์ง€ ๋ชจ๋ฅด๊ฒ ๋‹ค.. ๋ถ€๋“ค.. ๋‚ด๊ฐ€ ์ž˜๋ชปํ•œ๊ฑด์ง€, ๋…ผ๋ฌธ์—์„œ ๊ตฌ๋ผ์นœ๊ฑด์ง€ ๋ชจ๋ฅด๊ฒ ์ง€๋งŒ... ๊ผฌ๋ฆ„ํ•œ์ ์ด ํ•œ๋‘˜์ด ์•„๋‹ˆ๋‹ค. ์ผ๋‹จ ์„œ๋ก ์€ ์—ฌ๊ธฐ๊นŒ์ง€ํ•˜๊ณ  Abstract 1...
[๋…ผ๋ฌธ๋ฆฌ๋ทฐ]Are Prompt-Based Models Clueless?
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
Kavumba, Pride, Ryo Takahashi, and Yasuke Oda. "Are Prompt-based Models Clueless?." arXiv preprint arXiv:2205.09295 (2022). 2022๋…„๋„ 5์›”์— ACL์—์„œ ๋‚˜์˜จ ๋…ผ๋ฌธ์ž…๋‹ˆ๋‹ค. ์ฃผ์ œ๋Š” Prompt-based Models๋„ Superficial Cues๋ฅผ ์‚ฌ์šฉํ•˜๋Š”์ง€, ์‚ฌ์šฉํ•œ๋‹ค๋ฉด ์•…์šฉํ•˜์—ฌ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์ด ๋–จ์–ด์ง€๊ฒŒ ๋งŒ๋“ค์ง€๋Š” ์•Š๋Š”์ง€ ๋ถ„์„ํ•˜์—ฌ ํ™•์ธํ•˜๋Š” ๋…ผ๋ฌธ์ž…๋‹ˆ๋‹ค. (์ฐธ๊ณ ์ž๋ฃŒ) DiceLab SangHun Im
[DACON] ๋ฐ์ด์ฝ˜ ์‡ผํ•‘๋ชฐ ๋ฆฌ๋ทฐ ํ‰์  ๋ถ„๋ฅ˜ ๊ฒฝ์ง„๋Œ€ํšŒ
ยท
Artificial_Intelligence๐Ÿค–/Natural Language Processing
์ž์—ฐ์–ด ์ถ”๋ก ์ชฝ์„ ์—ฐ๊ตฌํ•˜๋ฉด์„œ, ์ง„ํ–‰์ค‘์ธ ์‹คํ—˜์ด ๋ง‰ํ˜€์„œ ์ž ๊น ํ™˜๊ธฐ์‹œํ‚ฌ๊ฒธ ์ž์—ฐ์–ด ๋Œ€ํšŒ๋ฅผ ์ฐพ์•„๋ดค์—ˆ๋‹ค. ์บ๊ธ€์—์„œ ์ง„ํ–‰์ค‘์ธ XNLI ๋Œ€ํšŒ, ํ•œ๊ตญ DACON์—์„œ ์ง„ํ–‰ํ•˜๋Š” ํ‰์  ๋ถ„๋ฅ˜ ๋Œ€ํšŒ ๋‘๊ฐ€์ง€๋ฅผ ์ฐพ์•„์„œ ์ด๋ฅผ ์ง„ํ–‰ํ•ด ๋ณด์•˜๋‹ค. ์ผ๋‹จ ์ด ๋Œ€ํšŒ๋ฅผ ์ฐพ์•˜์„ ๋‹น์‹œ์— ๋งˆ๊ฐ์ด D-1์ด๋ผ ๋‹ค๋ฅธ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์ง€๋Š” ๋ชปํ–ˆ๊ณ , ๋น ๋ฅด๊ฒŒ ์ œ์ถœํ•˜๊ธฐ ์œ„ํ•œ ๋ฒ ์ด์Šค๋ชจ๋ธ๋งŒ ์‚ฌ์šฉํ•˜์—ฌ ๋Œ€ํšŒ์— ์ฐธ๊ฐ€ํ•˜์˜€๋‹ค. ๋Œ€ํšŒ ์ฐธ์—ฌ ์ธ์›์€ 549๋ช… ์ •๋„์˜€๊ณ , 1๋“ฑ์˜ ์ ์ˆ˜๋Š” 0.71312 ์˜€๋‹ค. ๊ทธ๋ฆฌ๊ณ  ํ•˜๋ฃจ ํˆฌ์žํ•ด์„œ ์ œ์ถœํ•œ ๋‚ด ์ ์ˆ˜๋Š” 0.68888๋กœ, 1๋“ฑ๊ณผ์˜ ์ •ํ™•๋„ ์ฐจ์ด๊ฐ€ 0.02432(์•ฝ 2.4%) ์ฐจ์ด์˜€๋‹ค. ๋ฒ ์ด์ง ๋Œ€ํšŒ์ธ๋ฐ ํƒ‘10๋„ ๋ชปํ•ด์„œ ํ˜„ํƒ€๊ฐ€ ์˜ค๊ธด ํ–ˆ์ง€๋งŒ, ๊ทธ๋ž˜๋„ ๋น ๋ฅด๊ฒŒ ์ œ์ถœํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฑฐ์— ์˜์˜๋ฅผ ๋‘์—ˆ๋‹ค. ์ฒ˜์Œ์—๋Š” ์ผ๋‹จ ํ•œ๊ตญ์–ด ์ปค์Šคํ…€ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ๋งŒ๋“ค์–ด์„œ ์‚ฌ์šฉํ•˜๊ธฐ..
Liky
'Artificial_Intelligence๐Ÿค–' ์นดํ…Œ๊ณ ๋ฆฌ์˜ ๊ธ€ ๋ชฉ๋ก (2 Page)