kf-deberta-base / README.md
Arloha's picture
initial commit
363b171
|
raw
history blame
8.7 kB
metadata
license: mit
language:
  - ko
pipeline_tag: fill-mask

KF-DeBERTa

์นด์นด์˜ค๋ฑ…ํฌ & ์—ํ”„์—”๊ฐ€์ด๋“œ์—์„œ ํ•™์Šตํ•œ ๊ธˆ์œต ๋„๋ฉ”์ธ ํŠนํ™” ์–ธ์–ด๋ชจ๋ธ์„ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.

Model description

  • KF-DeBERTa๋Š” ๋ฒ”์šฉ ๋„๋ฉ”์ธ ๋ง๋ญ‰์น˜์™€ ๊ธˆ์œต ๋„๋ฉ”์ธ ๋ง๋ญ‰์น˜๋ฅผ ํ•จ๊ป˜ ํ•™์Šตํ•œ ์–ธ์–ด๋ชจ๋ธ ์ž…๋‹ˆ๋‹ค.
  • ๋ชจ๋ธ ์•„ํ‚คํ…์ณ๋Š” DeBERTa-v2๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šตํ•˜์˜€์Šต๋‹ˆ๋‹ค.
    • ELECTRA์˜ RTD๋ฅผ training objective๋กœ ์‚ฌ์šฉํ•œ DeBERTa-v3๋Š” ์ผ๋ถ€ task(KLUE-RE, WoS, Retrieval)์—์„œ ์ƒ๋‹นํžˆ ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜์—ฌ ์ตœ์ข… ์•„ํ‚คํ…์ณ๋Š” DeBERTa-v2๋กœ ๊ฒฐ์ •ํ•˜์˜€์Šต๋‹ˆ๋‹ค.
  • ๋ฒ”์šฉ ๋„๋ฉ”์ธ ๋ฐ ๊ธˆ์œต ๋„๋ฉ”์ธ downstream task์—์„œ ๋ชจ๋‘ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜์˜€์Šต๋‹ˆ๋‹ค.
    • ๊ธˆ์œต ๋„๋ฉ”์ธ downstream task์˜ ์ฒ ์ €ํ•œ ์„ฑ๋Šฅ๊ฒ€์ฆ์„ ์œ„ํ•ด ๋‹ค์–‘ํ•œ ๋ฐ์ดํ„ฐ์…‹์„ ํ†ตํ•ด ๊ฒ€์ฆ์„ ์ˆ˜ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค.
    • ๋ฒ”์šฉ ๋„๋ฉ”์ธ ๋ฐ ๊ธˆ์œต ๋„๋ฉ”์ธ์—์„œ ๊ธฐ์กด ์–ธ์–ด๋ชจ๋ธ๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์คฌ์œผ๋ฉฐ ํŠนํžˆ KLUE Benchmark์—์„œ๋Š” RoBERTa-Large๋ณด๋‹ค ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ํ™•์ธํ•˜์˜€์Šต๋‹ˆ๋‹ค.

Usage

from transformers import AutoModel, AutoTokenizer

model = AutoModel.from_pretrained("kakaobank/kf-deberta-base")
tokenizer = AutoTokenizer.from_pretrained("kakaobank/kf-deberta-base")

text = "์นด์นด์˜ค๋ฑ…ํฌ์™€ ์—ํ”„์—”๊ฐ€์ด๋“œ๊ฐ€ ๊ธˆ์œตํŠนํ™” ์–ธ์–ด๋ชจ๋ธ์„ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค."
tokens = tokenizer.tokenize(text)
print(tokens)

inputs = tokenizer(text, return_tensors="pt")
model_output = model(**inputs)
print(model_output)

Benchmark

  • ๋ชจ๋“  task๋Š” ์•„๋ž˜์™€ ๊ฐ™์€ ๊ธฐ๋ณธ์ ์ธ hyperparameter search๋งŒ ์ˆ˜ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค.
    • batch size: {16, 32}
    • learning_rate: {1e-5, 3e-5, 5e-5}
    • weight_decay: {0, 0.01}
    • warmup_proportion: {0, 0.1}

KLUE Benchmark

Model YNAT KLUE-ST KLUE-NLI KLUE-NER KLUE-RE KLUE-DP KLUE-MRC WoS AVG
F1 Pearsonr/F1 ACC F1-Entity/F1-Char F1-micro/AUC UAS/LAS EM/ROUGE JGA/F1-S
mBERT (Base) 82.64 82.97/75.93 72.90 75.56/88.81 58.39/56.41 88.53/86.04 49.96/55.57 35.27/88.60 71.26
XLM-R (Base) 84.52 88.88/81.20 78.23 80.48/92.14 57.62/57.05 93.12/87.23 26.76/53.36 41.54/89.81 72.28
XLM-R (Large) 87.30 93.08/87.17 86.40 82.18/93.20 58.75/63.53 92.87/87.82 35.23/66.55 42.44/89.88 76.17
KR-BERT (Base) 85.36 87.50/77.92 77.10 74.97/90.46 62.83/65.42 92.87/87.13 48.95/58.38 45.60/90.82 74.67
KoELECTRA (Base) 85.99 93.14/85.89 86.87 86.06/92.75 62.67/57.46 90.93/87.07 59.54/65.64 39.83/88.91 77.34
KLUE-BERT (Base) 86.95 91.01/83.44 79.87 83.71/91.17 65.58/68.11 93.07/87.25 62.42/68.15 46.72/91.59 78.50
KLUE-RoBERTa (Small) 85.95 91.70/85.42 81.00 83.55/91.20 61.26/60.89 93.47/87.50 58.25/63.56 46.65/91.50 77.28
KLUE-RoBERTa (Base) 86.19 92.91/86.78 86.30 83.81/91.09 66.73/68.11 93.75/87.77 69.56/74.64 47.41/91.60 80.48
KLUE-RoBERTa (Large) 85.88 93.20/86.13 89.50 84.54/91.45 71.06/73.33 93.84/87.93 75.26/80.30 49.39/92.19 82.43
KF-DeBERTa (Base) 87.51 93.24/87.73 88.37 89.17/93.30 69.70/75.07 94.05/87.97 72.59/78.08 50.21/92.59 82.83
  • ๊ตต์€๊ธ€์”จ๋Š” ๋ชจ๋“  ๋ชจ๋ธ์ค‘ ๊ฐ€์žฅ๋†’์€ ์ ์ˆ˜์ด๋ฉฐ, ๋ฐ‘์ค„์€ base ๋ชจ๋ธ ์ค‘ ๊ฐ€์žฅ ๋†’์€ ์ ์ˆ˜์ž…๋‹ˆ๋‹ค.

๊ธˆ์œต๋„๋ฉ”์ธ ๋ฒค์น˜๋งˆํฌ

Model FN-Sentiment (v1) FN-Sentiment (v2) FN-Adnews FN-NER KorFPB KorFiQA-SA KorHeadline Avg (FiQA-SA ์ œ์™ธ)
ACC ACC ACC F1-micro ACC MSE Mean F1
KLUE-RoBERTa (Base) 98.26 91.21 96.34 90.31 90.97 0.0589 81.11 94.03
KoELECTRA (Base) 98.26 90.56 96.98 89.81 92.36 0.0652 80.69 93.90
KF-DeBERTa (Base) 99.36 92.29 97.63 91.80 93.47 0.0553 82.12 95.27
  • FN-Sentiment: ๊ธˆ์œต๋„๋ฉ”์ธ ๊ฐ์„ฑ๋ถ„์„
  • FN-Adnews: ๊ธˆ์œต๋„๋ฉ”์ธ ๊ด‘๊ณ ์„ฑ๊ธฐ์‚ฌ ๋ถ„๋ฅ˜
  • FN-NER: ๊ธˆ์œต๋„๋ฉ”์ธ ๊ฐœ์ฒด๋ช…์ธ์‹
  • KorFPB: FinancialPhraseBank ๋ฒˆ์—ญ๋ฐ์ดํ„ฐ
    • Cite: Malo, Pekka, et al. "Good debt or bad debt: Detecting semantic orientations in economic texts." Journal of the Association for Information Science and Technology 65.4 (2014): 782-796.
  • KorFiQA-SA: FiQA-SA ๋ฒˆ์—ญ๋ฐ์ดํ„ฐ
    • Cite: Maia, Macedo & Handschuh, Siegfried & Freitas, Andre & Davis, Brian & McDermott, Ross & Zarrouk, Manel & Balahur, Alexandra. (2018). WWW'18 Open Challenge: Financial Opinion Mining and Question Answering. WWW '18: Companion Proceedings of the The Web Conference 2018. 1941-1942. 10.1145/3184558.3192301.
  • KorHeadline: Gold Commodity News and Dimensions ๋ฒˆ์—ญ๋ฐ์ดํ„ฐ
    • Cite: Sinha, A., & Khandait, T. (2021, April). Impact of News on the Commodity Market: Dataset and Results. In Future of Information and Communication Conference (pp. 589-601). Springer, Cham.

๋ฒ”์šฉ๋„๋ฉ”์ธ ๋ฒค์น˜๋งˆํฌ

Model NSMC PAWS KorNLI KorSTS KorQuAD Avg (KorQuAD ์ œ์™ธ)
ACC ACC ACC spearman EM/F1
KLUE-RoBERTa (Base) 90.47 84.79 81.65 84.40 86.34/94.40 85.33
KoELECTRA (Base) 90.63 84.45 82.24 85.53 84.83/93.45 85.71
KF-DeBERTa (Base) 91.36 86.14 84.54 85.99 86.60/95.07 87.01

License

KF-DeBERTa์˜ ์†Œ์Šค์ฝ”๋“œ ๋ฐ ๋ชจ๋ธ์€ MIT ๋ผ์ด์„ ์Šค ํ•˜์— ๊ณต๊ฐœ๋˜์–ด ์žˆ์Šต๋‹ˆ๋‹ค.
๋ผ์ด์„ ์Šค ์ „๋ฌธ์€ MIT ํŒŒ์ผ์—์„œ ํ™•์ธํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
๋ชจ๋ธ์˜ ์‚ฌ์šฉ์œผ๋กœ ์ธํ•ด ๋ฐœ์ƒํ•œ ์–ด๋– ํ•œ ์†ํ•ด์— ๋Œ€ํ•ด์„œ๋„ ๋‹น์‚ฌ๋Š” ์ฑ…์ž„์„ ์ง€์ง€ ์•Š์Šต๋‹ˆ๋‹ค.

Citation

@proceedings{jeon-etal-2023-kfdeberta,
  title         = {KF-DeBERTa: Financial Domain-specific Pre-trained Language Model},
  author        = {Eunkwang Jeon, Jungdae Kim, Minsang Song, and Joohyun Ryu},
  booktitle     = {Proceedings of the 35th Annual Conference on Human and Cognitive Language Technology},
  moth          = {oct},
  year          = {2023},
  publisher     = {Korean Institute of Information Scientists and Engineers},
  url           = {http://www.hclt.kr/symp/?lnb=conference},
  pages         = {143--148},
}