RefSum (Liu et al., 2021, NAACL)

๊น€์ˆ˜๋นˆยท2021๋…„ 11์›” 19์ผ
2

๋…ผ๋ฌธ ๋ฆฌ๋ทฐ

๋ชฉ๋ก ๋ณด๊ธฐ
2/14
post-thumbnail

๐Ÿ“„Paper: RefSum: Refactoring Neural Summarization

๐Ÿ’ก ๊ธฐ์กด 2-Stage Learning์˜ ํ•œ๊ณ„๋ฅผ ์™„ํ™”ํ•œ ํŒจ๋Ÿฌ๋‹ค์ž„ ์ œ์•ˆ

stage ๊ฐ„์˜ parameter ๊ณต์œ  & pretrain-then-finetune โ†’ ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์„ ์ƒํ˜ธ ๋ณด์™„์ ์œผ๋กœ ํ™œ์šฉํ•˜๋Š” ๋ฐฉ๋ฒ•


2-Stage learning์—๋Š” ์•„๋ž˜ 2๊ฐ€์ง€ ํ˜•ํƒœ๋กœ ๋‚˜๋‰œ๋‹ค.

  • stacking : ์„œ๋กœ ๋‹ค๋ฅธ ์—ฌ๋Ÿฌ base ๋ชจ๋ธ๋“ค๋กœ๋ถ€ํ„ฐ ์ƒ์„ฑ๋œ candidate summary๋“ค์„ ํ†ตํ•ด meta ๋ชจ๋ธ์ด ์ตœ์ข… summary๋ฅผ ์ƒ์„ฑ
  • re-ranking : ํ•˜๋‚˜์˜ base ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์ƒ์„ฑ๋œ ์—ฌ๋Ÿฌ candidate summary๋“ค์„ ํ†ตํ•ด meta ๋ชจ๋ธ์ด ์ตœ์ข… summary ์ƒ์„ฑ

โœจ Contribution ์ •๋ฆฌ

1. ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์„ ์‚ฌ์šฉํ•œ 2-stage learning์˜ ํ•œ๊ณ„ ๋ถ„์„


๊ธฐ์กด 2-stage learning์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ•œ๊ณ„๋ฅผ ๊ฐ€์ง€๋ฉฐ, ์ด๋กœ ์ธํ•ด ์„œ๋กœ ๋‹ค๋ฅธ ๋ชจ๋ธ์„ ์˜จ์ „ํžˆ ํ™œ์šฉํ•˜๊ธฐ ์–ด๋ ค์›€
  • Base-Meta Learning Gap Base model๊ณผ Meta model ๊ฐ„ parameter sharing์˜ ๋ถ€์žฌ๋กœ ์ธํ•ด ๋ฐœ์ƒ โ†’ Meta model์€ Base model์˜ output์„ ์˜จ์ „ํžˆ ํ™œ์šฉํ•˜์ง€ ๋ชปํ•จ
  • Train-Test Distribution Gap ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ Meta model์˜ output distribution๊ณผ ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ output distribution์€ ์ฐจ์ด๊ฐ€ ์žˆ์Œ โ†’ ํ›ˆ๋ จ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ output distribution์ด ๋” ์ •ํ™•ํ•จ

2. ๋‘ Gap์„ ์™„ํ™”ํ•œ ํŒจ๋Ÿฌ๋‹ค์ž„ ์ œ์•ˆ

๐Ÿ‘‰ Refactor

  • Base model์ด๋ฉด์„œ, Meta model๋กœ๋„ ํ™œ์šฉํ•  ์ˆ˜ ์žˆ์–ด Parameter sharing์ด ๊ฐ€๋Šฅํ•จ
    โ†’ Base-Meta Gap ์™„ํ™”
  • 2๋ฒˆ์˜ ํ•™์Šต์œผ๋กœ ๋‹ค์–‘ํ•œ candidate summary๋ฅผ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ์Œ
    โ†’ Train-Test Gap ์™„ํ™”

โ‘  pre-train : Input document๋กœ๋ถ€ํ„ฐ candidate summary ์ƒ์„ฑ
โ‘ก fine-tune : Base model์˜ ๋‹ค์–‘ํ•œ output์œผ๋กœ๋ถ€ํ„ฐ new candidate summary ์ƒ์„ฑ


Refactor๋Š” Base model๊ณผ Meta model์„ ๋ถ„๋ฆฌํ•˜์ง€ ์•Š๊ณ  ๊ณตํ†ต์œผ๋กœ ์‚ฌ์šฉํ•จ์œผ๋กœ์จ parameter sharing์ด ๊ฐ€๋Šฅํ•จ

๊ธฐ์กด์˜ 2 stage learning์˜ ๋ถ„๋ฆฌ๋œ base-meta ๋ชจ๋ธ

C=BASE(D,T,S,ฮธbase)\begin{aligned} C = BASE(D, \mathcal{T}, S, \theta^{base}) \end{aligned}
Cโˆ—=META(D,C,ฮธmeta)\begin{aligned} C^{*} = META(D, \mathcal{C}, \theta^{meta}) \end{aligned}

Refactor๋ฅผ ํ†ตํ•ด ํ†ตํ•ฉ๋œ base-meta model

Cโˆ—=REFACTOR(D,C,ฮธrefactor)C^{*} = REFACTOR(D,\mathcal{C}, \theta^{refactor}) \\

์•„๋ž˜๋Š” Liu et al.์˜ ์‹คํ—˜์„ ์š”์•ฝํ•œ ๋‚ด์šฉ์ด๋‹ค.

Baseย : ๋ฒ ์ด์Šค ๋ชจ๋ธ๋งŒ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ
Pre-trainedย : ์‚ฌ์ „ ํ•™์Šต๋œ Refactor๋ฅผ ๋ฉ”ํƒ€ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ
Supervisedย : Refactor๋ฅผ ๋ฉ”ํƒ€ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๋ฉฐ
ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ output์œผ๋กœ๋งŒ ํ•™์Šตํ•œ ๊ฒฝ์šฐ (์‚ฌ์ „ ํ•™์Šต์„ ๊ฑฐ์น˜์ง€ ์•Š์€ Fine-tuned)
Fine-tunedย : ์‚ฌ์ „ ํ•™์Šต๋œ Refactor์— ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ output์œผ๋กœ
ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย ย fine-tuningํ•œ Refactor๋ฅผ ๋ฉ”ํƒ€ ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ
โ€ : ๋ฒ ์ด์Šค ๋ชจ๋ธ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ๋›ฐ์–ด๋‚œ ๊ฒฝ์šฐ


1) Single System re-ranking (on CNN/DM)

๐Ÿ’ก ์‹คํ—˜ ๊ฒฐ๊ณผ 1

  1. Refactor๋Š” ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค.
  2. Fine-tuned Refactor๋Š” Supervised Refactor์˜ ์„ฑ๋Šฅ์„ ๋›ฐ์–ด ๋„˜๋Š”๋‹ค.
    (์‚ฌ์ „ ํ•™์Šต์˜ ์ค‘์š”์„ฑ)

2) Multi System Stacking (on CNN/DM)

  • Summary-level combination
    ๊ฐ ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ output์„ summary-level๋กœ ์กฐํ•ฉํ•˜์—ฌ candidate summary set์œผ๋กœ ์‚ฌ์šฉ

Baseย : ํ•˜๋‚˜์˜ ๋ฒ ์ด์Šค ๋ชจ๋ธ๋งŒ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ
Twoย : ๋ฒ ์ด์Šค ๋ชจ๋ธ์„ 2๊ฐœ (BART, ์‚ฌ์ „ ํ•™์Šต๋œ Refactor) ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ
Threeย : ๋ฒ ์ด์Šค ๋ชจ๋ธ์„ 3๊ฐœ (BART, ์‚ฌ์ „ ํ•™์Šต๋œ Refactor, GSum) ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ

  • Sentence-level combination
    ๊ฐ ๋ฒ ์ด์Šค ๋ชจ๋ธ(BART, ์‚ฌ์ „ ํ•™์Šต๋œ Refactor)์˜ output์„ sentence-level๋กœ ์กฐํ•ฉํ•˜์—ฌ candidate summary set์œผ๋กœ ์‚ฌ์šฉ

๐Ÿ’ก ์‹คํ—˜ ๊ฒฐ๊ณผ 2

  1. ์‚ฌ์ „ ํ•™์Šต๋œ Refactor๋Š” ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ๋›ฐ์–ด๋„˜๋Š”๋‹ค.
  2. fine-tuning์ด ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค.

โ—๏ธ ์˜ˆ์™ธ

  1. Sentence-level combination์—์„œ Supervised Refactor๊ฐ€
    Fine-tuned Refactor์™€ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ž„

    pre-training data์— ๋น„ํ•ด fine-tuning data๊ฐ€ ์ƒ๋‹นํžˆ ๋งŽ๊ธฐ ๋•Œ๋ฌธ

  2. Summary-level combination์—์„œ ๋ฒ ์ด์Šค ๋ชจ๋ธ์ธ GSum์ด
    Pre-trained Refactor (Three) ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ž„

    GSum์ด ๋‹ค๋ฅธ ๋‘ ๋ฒ ์ด์Šค ๋ชจ๋ธ๋ณด๋‹ค ๋›ฐ์–ด๋‚œ ์„ฑ๋Šฅ์„ ๋ณด์ด๊ธฐ ๋•Œ๋ฌธ

3) Generalization on 19 Top-performing System (on CNN/DM)

Base ๋ชจ๋ธ : 19๊ฐœ์˜ Top-performing system
Meta ๋ชจ๋ธ : pre-trained Refactor

์ด ์‹คํ—˜์—์„œ Meta ๋ชจ๋ธ์ธ pretrained Refactor๋Š” fine-tuning ์—†์ด Base ๋ชจ๋ธ์˜ output์„ ํ†ตํ•ด summary๋ฅผ ์ถ”์ถœํ•˜์˜€๋‹ค.

19๊ฐœ์˜ ๋ชจ๋ธ์„ ์กฐํ•ฉํ•˜๋ฉฐ multi-system stacking ์‹คํ—˜์„ ์ˆ˜ํ–‰ํ•˜์˜€๋‹ค.

bin : ROUGE score ๋ฒ”์œ„
#sys : ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ๊ฐœ์ˆ˜
Ours : pre-trained Refactor

x์ถ•์€ ๋ฒ ์ด์Šค ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋œ ์‹œ์Šคํ…œ๋“ค ๊ฐ„์˜ ์„ฑ๋Šฅ ์ฐจ์ด์ด๋ฉฐ, ROUGE-1 score๋ฅผ ํ†ตํ•ด ์ธก์ •๋˜์—ˆ๋‹ค.
y์ถ•์€ Refactor๋ฅผ ์‚ฌ์šฉํ•˜์˜€์„ ๋•Œ Single base model์„ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ ๋Œ€๋น„ ๋ชจ๋ธ ์„ฑ๋Šฅ ํ–ฅ์ƒ๋ฅ ์„ ์˜๋ฏธํ•œ๋‹ค.

๐Ÿ’ก ์‹คํ—˜ ๊ฒฐ๊ณผ 3

  1. ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๊ฐ€์ง„ ๋ชจ๋ธ๋“ค์ด ๋ฒ ์ด์Šค ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์„ ๋•Œ ์ข‹์€ ์„ฑ๋Šฅ์„ ๋ณด์˜€๋‹ค.
  2. ๋ฒ ์ด์Šค ๋ชจ๋ธ ๊ฐ„ ์„ฑ๋Šฅ ์ฐจ์ด๊ฐ€ ์ ์„ ์ˆ˜๋ก Refactor๋ฅผ ์ถ”๊ฐ€ํ•œ ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋œ๋‹ค.

๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์…‹์—์„œ Re-ranking ์‹คํ—˜ ์ˆ˜ํ–‰ (on XSum, PubMed, WikiHow)

โœ๏ธ ์‹คํ—˜

Refactor ์‚ฌ์ „ ํ•™์Šต์€ pre-trained BART๋ฅผ ํ†ตํ•ด ์ˆ˜ํ–‰
XSum์—์„œ๋Š” PEGASUS ๋ชจ๋ธ์„, ๋‹ค๋ฅธ ๋ฐ์ดํ„ฐ์…‹์—๋Š” BART๋ฅผ ๋ฒ ์ด์Šค ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ

๐Ÿ’ก ์‹คํ—˜ ๊ฒฐ๊ณผ 4

  1. Refactor๋Š” ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚จ๋‹ค.
  2. XSum์˜ ๊ฒฝ์šฐ, Pre-trained Refactor๊ฐ€ Fine-tuned Refactor์˜ ์„ฑ๋Šฅ์„ ๋›ฐ์–ด ๋„˜์—ˆ๋‹ค.

    BART๋ฅผ ํ†ตํ•ด ์–ป์€ pre-training data๊ฐ€ PEGASUS output์„ re-rankingํ•˜๊ธฐ ์œ„ํ•œ Refactor๋ฅผ ํ•™์Šต์‹œํ‚ค๊ธฐ์— ์ถฉ๋ถ„ํ•˜๊ธฐ ๋•Œ๋ฌธ.


5) Fine-grained Analysis

๋ชจ๋ธ ํ”„๋ ˆ์ž„์›Œํฌ
:์‹คํ—˜ 2์˜ Multi System Stacking: Summary-level combination (Two)

๋ฒ ์ด์Šค ๋ชจ๋ธ์ธ BART์™€ pre-trained Refactor๊ฐ€ ์ƒ์„ฑํ•œ candidate summary์˜ ROUGE score ์ฐจ์ด Performance Gap)์— ๋”ฐ๋ฅธ ๋ฉ”ํƒ€ ๋ชจ๋ธ์˜ ์ •ํ™•๋„ ์ธก์ •

x ์ถ•์€ ๋‘ ๋ฒ ์ด์Šค ๋ชจ๋ธ์˜ ROUGE score ์ฐจ์ด๋ฅผ,
y์ถ•์€ ๋ฉ”ํƒ€ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๋ฅผ ์˜๋ฏธํ•œ๋‹ค.

๐Ÿ’ก ์‹คํ—˜ ๊ฒฐ๊ณผ 5

  1. Performance Gap์ด ํด์ˆ˜๋ก ๋ฉ”ํƒ€ ๋ชจ๋ธ์˜ ์ •ํ™•๋„๊ฐ€ ํ–ฅ์ƒ๋จ

    ๋‘ ๋ฒ ์ด์Šค ๋ชจ๋ธ์ด ์„œ๋กœ ์ƒํ˜ธ ๋ณด์™„์ ์œผ๋กœ ๋™์ž‘ํ•˜๊ณ  ์žˆ์Œ์„ ์•Œ ์ˆ˜ ์žˆ์Œ

Performance Gap์ด ํฌ๋‹ค๋Š” ๊ฒƒ์€ ์–ด๋Š ํ•œ ๋ชจ๋ธ์˜ ROUGE score๊ฐ€ ๋‚ฎ๊ณ  ๋‹ค๋ฅธ ๋ชจ๋ธ์˜ ROUGE score๊ฐ€ ๋†’๋‹ค๋Š” ๊ฒƒ์ธ๋ฐ, ์ด๋ฅผ ํ†ตํ•ด ๋‘ ๋ฒ ์ด์Šค ๋ชจ๋ธ์ด ์„œ๋กœ ์ƒํ˜ธ ๋ณด์™„์ ์œผ๋กœ ๋™์ž‘ํ•˜๊ณ  ์žˆ์Œ์„ ์ง์ž‘ํ•  ์ˆ˜ ์žˆ๋‹ค.

(์–ด๋Š ํ•œ ๋ชจ๋ธ์˜ ROUGE score๊ฐ€ ๊ณ„์† ๋‚ฎ๊ณ , ๋‹ค๋ฅธ ๋ชจ๋ธ์ด ๋†’์€ score๋ฅผ ๊ฐ€์ง€์ง€ ์•Š๋‚˜ ์ƒ๊ฐํ•ด๋ณผ ์ˆ˜ ์žˆ๊ฒ ์ง€๋งŒ, ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ ๊ฒƒ์€ ๋น„์Šทํ•œ ์„ฑ๋Šฅ์„ ๊ฐ€์ง„ ๋ชจ๋ธ๋“ค์„ ๋ฒ ์ด์Šค ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ์ด๊ธฐ ๋•Œ๋ฌธ์— ์ƒํ˜ธ ๋ณด์™„์ ์œผ๋กœ ๋™์ž‘ํ•  ๊ฐ€๋Šฅ์„ฑ์ด ๋†’๋‹ค)



๋ฆฌ๋ทฐ

๋…ผ๋ฌธ์„ ์ฝ๊ณ  ์ดํ•ด๊ฐ€ ๊ฐ€์ง€ ์•Š๋Š” ๋ถ€๋ถ„์ด ์žˆ๋Š”๋ฐ,

๋…ผ๋ฌธ์—์„œ Base-Meta ๊ฐ„ Gap์ด 2 stage learning ์„ฑ๋Šฅ์— ์˜ํ–ฅ์„ ์ฃผ๋Š” ์ฃผ์š” ์š”์ธ์ด๋ผ๊ณ  ์–ธ๊ธ‰ํ•˜์˜€์œผ๋ฉฐ, ์ด gap์„ ์™„ํ™”ํ•  REFACTOR๋ฅผ ๋„์ž…ํ•˜์˜€๋‹ค.

ํ•˜์ง€๋งŒ, ์‹คํ—˜ ๋‚ด์šฉ์„ ๋ณด๋ฉด Base-Meta ๊ฐ„ Gap์€ ํ•ด๊ฒฐ๋˜์ง€ ์•Š์€ ๊ฒƒ์œผ๋กœ ์ถ”์ •๋œ๋‹ค.

Train-Test gap์€ Fine-tuning์„ ํ†ตํ•ด ํ•ด์†Œ๋  ์ˆ˜ ์žˆ์ง€๋งŒ, Base-Meta ๊ฐ„ Gap์€ Parameter sharing์„ ํ†ตํ•ด ํ•ด์†Œ๋˜์–ด์•ผ ํ•˜๋Š”๋ฐ (๋…ผ๋ฌธ์—์„œ ์ฃผ์žฅํ•œ ๋ฐ”์— ์˜ํ•˜๋ฉด),

์‹คํ—˜ ๋‚ด์šฉ์„ ๋ณด๋ฉด Reranking task์—์„œ๋Š” Refactor๋Š” Meta ๋ชจ๋ธ๋กœ๋งŒ ์‚ฌ์šฉ๋˜์—ˆ์œผ๋ฉฐ, Multi-stacking task์ธ ์‹คํ—˜ 2์—์„œ Base ๋ชจ๋ธ๋กœ ์‚ฌ์šฉ๋˜์—ˆ์ง€๋งŒ, ์ด๋•Œ Base ๋ชจ๋ธ์€ Refactor ์™ธ์—๋„ GSum, BART๋„ ์‚ฌ์šฉ๋˜์—ˆ๋‹ค. Refactor๊ฐ€ ์•„๋‹Œ ๋‹ค๋ฅธ Base ๋ชจ๋ธ๊ณผ Meta ๋ชจ๋ธ์€ Parameter Sharing์ด ์ด๋ค„์ง€์ง€ ์•Š๋Š” ๊ฒƒ์œผ๋กœ ๋ณด์ด๋ฉฐ, BART์™€ GSum์˜ output์„ Meta ๋ชจ๋ธ์ด ์˜จ์ „ํžˆ ์ด์šฉํ•œ ๊ฒƒ์ธ์ง€ ์•Œ ์ˆ˜ ์—†๋‹ค.

๋”ฐ๋ผ์„œ ๋…ผ๋ฌธ์˜ ์ฃผ์š” contribution ์ค‘ ํ•˜๋‚˜์ธ Parameter Sharing์ด ๋ชจ๋ธ ์„ฑ๋Šฅ๊ณผ ์—ฐ๊ด€์ด ์žˆ๋‹ค๋Š” ์‹คํ—˜์€ ์—†๋‹ค.


Base-Meta๊ฐ„ gap์„ ํ•ด์†Œํ•˜๋Š” Parameter sharing์ด ์„ฑ๋Šฅ๊ณผ ์–ด๋–ค ์—ฐ๊ด€์ด ์žˆ๋Š”์ง€ ๋ฐํžˆ๋Š” ์‹คํ—˜์ด ์—†์—ˆ๊ธฐ ๋•Œ๋ฌธ์—,
๋…ผ๋ฌธ์˜ ์ €์ž์ด์‹  Yixin Liu๋ถ„๊ป˜ ์ด contribution์— ๋Œ€ํ•œ ์งˆ๋ฌธ์„ ๋“œ๋ ธ๋‹ค.



๋…ผ๋ฌธ์˜ ์ €์ž๋กœ๋ถ€ํ„ฐ ์ด์— ๋Œ€ํ•œ ๋‹ต๋ณ€์„ ๋ฐ›์•˜๋‹ค.
Re-ranking task์˜ candidates๋Š” ์ถ”์ƒ ์š”์•ฝ ๋ชจ๋ธ๋กœ๋ถ€ํ„ฐ ์ƒ์„ฑ๋œ ์ถ”์ƒ ์š”์•ฝ๋ฌธ์œผ๋กœ ๊ตฌ์„ฑํ–ˆ๊ธฐ ๋•Œ๋ฌธ์— ์ถ”์ถœ ์š”์•ฝ ๋ชจ๋ธ์ธ Refactor๋ฅผ Base ๋ชจ๋ธ๋กœ ์‚ฌ์šฉํ•˜๊ธฐ ์ ์ ˆํ•˜์ง€ ์•Š์•˜๊ณ ,
Parameter sharing์€ ๋…ผ๋ฌธ์—์„œ ์ œ์•ˆํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์˜ ํŠน์ง• ์ค‘ ํ•˜๋‚˜๋กœ, ์„ฑ๋Šฅ ํ–ฅ์ƒ๊ณผ ์ง์ ‘์ ์ธ ์—ฐ๊ด€์€ ์—†์œผ๋‚˜, ํ•œ๋ฒˆ์˜ ํ•™์Šต์œผ๋กœ pre-trained Meta ๋ชจ๋ธ์„ ์–ป๋Š” ๋™์‹œ์— Base ๋ชจ๋ธ์„ ์–ป์„ ์ˆ˜ ์žˆ๋‹ค๋Š” ์žฅ์ ์ด ์žˆ๋‹ค.

profile
์—ด์‹ฌํžˆ ๋ฐฐ์šฐ๋Š” ๋‚ด๊ฐ€ ๋˜์ž

0๊ฐœ์˜ ๋Œ“๊ธ€