Notice
Recent Posts
Recent Comments
ยซ   2024/11   ยป
์ผ ์›” ํ™” ์ˆ˜ ๋ชฉ ๊ธˆ ํ† 
1 2
3 4 5 6 7 8 9
10 11 12 13 14 15 16
17 18 19 20 21 22 23
24 25 26 27 28 29 30
Tags more
Archives
Today
Total
๊ด€๋ฆฌ ๋ฉ”๋‰ด

๐ŸŒฒ์ž๋ผ๋‚˜๋Š”์ฒญ๋…„

[๋…ผ๋ฌธ๋ฆฌ๋ทฐ]Few Shot Dialogue State Tracking using Meta-learning(2021) ๋ณธ๋ฌธ

๋…ผ๋ฌธ๋ฆฌ๋ทฐ

[๋…ผ๋ฌธ๋ฆฌ๋ทฐ]Few Shot Dialogue State Tracking using Meta-learning(2021)

JihyunLee 2021. 8. 26. 16:12
๋ฐ˜์‘ํ˜•

์ œ๋ชฉ : Few Shot Dialogue State Tracking using Meta-learning

์ €์ž : Saket Dingliwal, Shuyang Gao, Sanchit Agarwal, Chien-Wei Lin, Tagyoung Chung, Dilek Hakkani-T¨ur

๋ฐœํ–‰๋…„๋„ : 2021

paper : https://arxiv.org/abs/2101.06779

code : https://github.com/saketdingliwal/Few-Shot-DST

Review

dialogue state tracking(DST)๊ฐ€ ํŠน์ • ๋ชฉ์ ์„ ๊ฐ€์ง„ chatbot์„ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด ๋งŽ์ด ์‚ฌ์šฉ๋˜๋Š” ๋ฐฉ๋ฒ•์ด์ง€๋งŒ, dst๋ฅผ ๋งŒ๋“ค๊ธฐ ์œ„ํ•ด์„œ๋Š” ์ž˜ annotation ๋œ ๋ฐ์ดํ„ฐ๊ฐ€ ํ•„์š”ํ•˜๋‹ค. ๊ทธ๋Ÿฌ๋‚˜ annotation์ด ์ธ๋ ฅ์ด ๋งŽ์ด ๋“ค์–ด๊ฐ€๋Š” ๋ฐฉ๋ฒ•์ด๊ธฐ์—, zero shot/few shot ์˜ ๋ฐฉ์‹์˜ dst๊ฐ€ ํ•„์ˆ˜์ ์ด๋‹ค. ์ด ๋…ผ๋ฌธ์€ D-REPTILE์ด๋ผ๋Š” meta learning ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๊ฐœ๋ฐœํ•˜์—ฌ, zeroshot/fewshot ํ™˜๊ฒฝ์—์„œ๋„ ์ž˜ ์ ์šฉ๋˜๋Š” meta-learning dst ์‹œ์Šคํ…œ์„ ๋งŒ๋“ค์—ˆ๋‹ค.

Multi Woz data ์— ์ ์šฉํ•œ ๊ฒฐ๊ณผ ์ด๋ฏธ์ง€ ์ค‘ ํ•˜๋‚˜. ํŒŒ๋ž€์ƒ‰ ์‹ค์„ ์ด ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ ๋ชจ๋ธ

๊ฒฐ๊ณผ๋ถ€ํ„ฐ ๋ณด๋ฉด x์ถ•์€ ํ•™์Šตํ•œ dialogue, ์„ธ๋กœ์ถ•์€ joint goal accuracy ๋ฅผ ๋‚˜ํƒ€๋‚ด๋Š”๋ฐ, zero/few shot ํ™˜๊ฒฝ์—์„œ๋„ ์„ฑ๋Šฅ์ด ์–ด๋Š์ •๋„ ์ž˜ ๋‚˜์˜ค๊ณ  ์žˆ๋Š”๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.

meta-learning?

๋”๋ณด๊ธฐ

learning to learn์ด๋ผ๊ณ  ํ•˜๋ฉฐ, ์ ์€ ๋ฐ์ดํ„ฐ๋กœ๋„ ์ผ๋ฐ˜ํ™”๊ฐ€ ๊ฐ€๋Šฅํ•˜๋„๋ก ํ•˜์ž๋Š” ๋ชฉ์ ์„ ๊ฐ€์ง€๊ณ  ์žˆ๋‹ค. 
๊ฐœ๋…์ด ํฌํ•จํ•˜๊ณ  ์žˆ๋Š” ๋‚ด์šฉ์ด ๋ฐฉ๋Œ€ํ•ด์„œ, ๋…ผ๋ฌธ์— ๋‚˜์˜จ ๋‚ด์šฉ๋งŒ ์ •๋ฆฌํ•˜์ž๋ฉด, ๋…ผ๋ฌธ์—์„œ๋Š” REPTILE(2018) ์ด๋ผ๋Š” few shot learning ์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ๋ณ€ํ˜•ํ•˜์—ฌ ์‚ฌ์šฉํ•˜์˜€๋‹ค. meta learning ์˜ ์—ฌ๋Ÿฌ ๋ฐฉ์‹ ์ค‘, Optimization ๋ฐฉ์‹์„ ๋ณ€ํ˜•ํ•ด์„œ few shot learning ์ด ๊ฐ€๋Šฅํ•˜๋„๋ก ํ•œ ๋…ผ๋ฌธ์ด๊ณ  ์ด์ „ ๋ฐฉ์‹(MAML - model agnostic meta learning)์— ๋น„ํ•ด ์—ฐ์‚ฐ๊ณผ ๋ฉ”๋ชจ๋ฆฌ ์‚ฌ์šฉ๋Ÿ‰์—์„œ ์ด์ ์ด ์žˆ๋Š” ๋ฐฉ๋ฒ•์ด๋‹ค. 

Reptile ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์„ค๋ช…. ๋…ผ๋ฌธ์—์„œ ๊ฐ€์ ธ์˜ด

์ž์„ธํžˆ ๋ณด๋ฉด ๋‚ด๊ฐ€ ์ดํ•ดํ•˜๊ธฐ๋ก , ์ „์ฒด task(data)๋ฅผ samplingํ•˜์—ฌ loss๋ฅผ ๊ตฌํ•ด ํ‰๊ท ์„ ๋‚ธ ๋’ค ์—…๋ฐ์ดํŠธ ํ•˜๋Š” ๋ฐฉ์‹์„ ์—ฌ๋Ÿฌ๋ฒˆ ์ ์šฉํ•˜๋Š” ๋“ฏ ํ•œ๋ฐ, ์ด๊ฒŒ ์–ด๋–ป๊ฒŒ ์ž˜ ๋˜๋Š”์ง€๋Š” ์˜๋ฌธ์ด๋‹ค. ๊ทธ๋ž˜๋„ ๊ฒฐ๊ณผ๋ฅผ ๋ณด๋ฉด ์ž˜ ๋œ๋‹ค๊ณ  ํ•˜๋Š” ์„ค๋ช…์ด ๋งŽ์•„์„œ, ๋‹ค์Œ์— ๋” ๊ณต๋ถ€๋ฅผ ํ•ด๋ด์•ผ๊ฒ ๋‹ค.

์•„๋ž˜ ๊ทธ๋ฆผ์€ MAML์— ์žˆ๋Š” ๊ทธ๋ฆผ์ด๋‹ค. Meta learning์€ ์•„๋ž˜์™€ ๊ฐ™์ด ์ ์€ ๋ฐ์ดํ„ฐ sample์— ๋Œ€ํ•ด loss๋ฅผ ์—ฌ๋Ÿฌ๊ฐœ ๊ตฌํ•ด์„œ ํ‰๊ท ๋‚ด์„œ ํ•œ๋ฒˆ์— ์ด๋™ํ•œ๋‹ค๊ณ  ์ดํ•ดํ–ˆ๋‹ค.

MAML ๋…ผ๋ฌธ์—์„œ ๊ฐ€์ ธ์˜จ ๊ทธ๋ฆผ

 

Methodology

๋ฉ”ํƒ€ ๋Ÿฌ๋‹์˜ ์•Œ๊ณ ๋ฆฌ์ฆ˜ ์ค‘ REPTILE์„ ๋ณ€๊ฒฝ์‹œํ‚จ D-REPTILE์•Œ๊ณ ๋ฆฌ์ฆ˜์„ ์‚ฌ์šฉํ•œ๋‹ค

์ถœ์ฒ˜ : Few Shot Dialogue State Tracking using Meta-learning

input์ธ D๋Š” domain์„ ์˜๋ฏธํ•˜๋ฉฐ paramter๋Š” meta๋ชจ๋ธ์„ ํ•™์Šต์‹œํ‚ค๋Š”๋ฐ ํ•„์š”ํ•œ ํ•˜์ดํผ ํŒŒ๋ผ๋ฏธํ„ฐ, ๊ทธ๋ฆฌ๊ณ  ๊ฒฐ๊ณผ ๊ฐ’์œผ๋กœ  ํ•™์Šต๋œ meta model์ด ๋‚˜์˜จ๋‹ค.

๊ณผ์ •์„ ๋ณด๋ฉด ๊ฐ domain(hotel.area, hotel.stars)์„ ํ•™์Šต์‹œ์ผœ ์–ป์€ ์ƒˆ๋กœ์šด state๋ฅผ ๊ธฐ์กด state์™€ ์ฐจ๋ฅผ ๊ตฌํ•˜๊ณ , ์ด๋ฅผ ๋„๋ฉ”์ธ๋ณ„๋กœ ํ‰๊ท ๋‚ธ ๊ฐ’์„ meta model์— ์ ์šฉํ•˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ๋‹ค.(๊ธฐ์กด REIPTILE๊ณผ ๊ฑฐ์˜ ๋™์ผ)

 

์ผ๋ถ€๋งŒ tagging๋œ ๋ฐ์ดํ„ฐ์—๋„ ์ ์šฉ์ด ๊ฐ€๋Šฅํ•˜๋‹ค๊ณ  ํ•˜๋‹ˆ, ์‹ค์ œ ํ˜„์‹ค ๋ฐ์ดํ„ฐ์— ์ ์šฉํ•ด์„œ ์‹คํ—˜ํ•ด ๋ณผ ๋งŒ ํ•˜๋‹ค๋Š” ์ƒ๊ฐ์ด ๋“ ๋‹ค.

reference

 

๋ฉ”ํƒ€๋Ÿฌ๋‹ : https://talkingaboutme.tistory.com/entry/DL-Meta-Learning-Learning-to-Learn-Fast๋…ผ๋ฌธ : https://arxiv.org/abs/2101.06779

 

[DL] Meta-Learning: Learning to Learn Fast

(ํ•ด๋‹น ๊ธ€์€ OpenAI Engineer์ธ Lilian Weng์˜ ํฌ์ŠคํŠธ ๋‚ด์šฉ์„ ์›์ €์ž ๋™์˜ํ•˜์— ๋ฒˆ์—ญํ•œ ๋‚ด์šฉ์ž…๋‹ˆ๋‹ค.) Meta-Learning: Learning to Learn Fast Meta-learning, also known as “learning to learn”, inte..

talkingaboutme.tistory.com

 

๋ฐ˜์‘ํ˜•