KoBERT ๊ธฐ๋ฐ˜ ํ•œ๊ตญ์–ด ๊ฐ์ • ๋ถ„๋ฅ˜ ๋ชจ๋ธ

์ด ํ”„๋กœ์ ํŠธ๋Š” ํ•œ๊ตญ์–ด ํ…์ŠคํŠธ์˜ ๊ฐ์ •์„ ๋ถ„๋ฅ˜ํ•˜๋Š” KoBERT ๊ธฐ๋ฐ˜์˜ ๊ฐ์ • ๋ถ„๋ฅ˜ ๋ชจ๋ธ์„ ํ•™์Šตํ•˜๊ณ  ํ™œ์šฉํ•˜๋Š” ์ฝ”๋“œ๋ฅผ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ชจ๋ธ์€ ์ž…๋ ฅ๋œ ํ…์ŠคํŠธ๊ฐ€ ๋ถ„๋…ธ(Anger), ๋‘๋ ค์›€(Fear), ๊ธฐ์จ(Happy), ํ‰์˜จ(Tender), ์Šฌํ””(Sad) ์ค‘ ์–ด๋–ค ๊ฐ์ •์— ํ•ด๋‹นํ•˜๋Š”์ง€๋ฅผ ์˜ˆ์ธกํ•ฉ๋‹ˆ๋‹ค.

1. ๋ชจ๋ธ ํ•™์Šต ๊ณผ์ •

Colab ํ™˜๊ฒฝ ์„ค์ • ๋ฐ ๋ฐ์ดํ„ฐ ์ค€๋น„

  1. ํ•„์š” ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ ์„ค์น˜: transformers, datasets, torch, pandas, scikit-learn ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ์„ค์น˜ํ•ฉ๋‹ˆ๋‹ค.

  2. ๋ฐ์ดํ„ฐ ๋ถˆ๋Ÿฌ์˜ค๊ธฐ: ai hub ์— ๋“ฑ๋ก๋œ ํ•œ๊ตญ์–ด ๊ฐ์„ฑ ๋Œ€ํ™” ๋ฐ์ดํ„ฐ๋กœ๋ถ€ํ„ฐ ๊ฐ์ • ๋ถ„๋ฅ˜์šฉ CSV ํŒŒ์ผ์„ ๋ถˆ๋Ÿฌ์˜ต๋‹ˆ๋‹ค.

  3. ๋ฐ์ดํ„ฐ์…‹ ์ค€๋น„:

    • ํ•™์Šต/๊ฒ€์ฆ ๋ฐ์ดํ„ฐ ๋ถ„ํ• : 80%๋Š” ํ•™์Šต ๋ฐ์ดํ„ฐ๋กœ, 20%๋Š” ๊ฒ€์ฆ ๋ฐ์ดํ„ฐ๋กœ ์‚ฌ์šฉ.
    • HuggingFace Dataset ํ˜•์‹ ๋ณ€ํ™˜: Pandas DataFrame์„ HuggingFace Dataset์œผ๋กœ ๋ณ€ํ™˜.
    • ๋ ˆ์ด๋ธ” ์ปฌ๋Ÿผ๋ช… ๋ณ€๊ฒฝ: ๊ฐ์ • ๋ ˆ์ด๋ธ”์„ ๋‚˜ํƒ€๋‚ด๋Š” label_int ์ปฌ๋Ÿผ์„ labels๋กœ ๋ณ€๊ฒฝ.
    • ๋ฐ์ดํ„ฐ ํ† ํฐํ™”: monologg/kobert ํ† ํฌ๋‚˜์ด์ €๋ฅผ ์ด์šฉํ•ด ์ž…๋ ฅ ํ…์ŠคํŠธ๋ฅผ ํ† ํฐํ™”.
    • ํ˜•์‹ ๋ณ€ํ™˜: input_ids, attention_mask, labels๋งŒ ๋‚จ๊ฒจ ํ•™์Šต ์ค€๋น„ ์™„๋ฃŒ.
  4. ๋ชจ๋ธ ๋ฐ ํ•™์Šต ์„ค์ •:

    • ๋ชจ๋ธ: monologg/kobert ๋ชจ๋ธ์„ ๋ถˆ๋Ÿฌ์™€ 5๊ฐœ์˜ ๊ฐ์ • ๋ ˆ์ด๋ธ”์„ ๋ถ„๋ฅ˜ํ•˜๋„๋ก ์„ค์ •.
    • ํ•™์Šต ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ:
      • learning_rate=2e-5, num_train_epochs=10, batch_size=16.
      • F1 ์Šค์ฝ”์–ด๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ๋ฒ ์ŠคํŠธ ๋ชจ๋ธ ์ €์žฅ.
      • Early stopping ์ ์šฉ.
  5. ํ•™์Šต ์ง„ํ–‰ ๋ฐ ๋ชจ๋ธ ์ €์žฅ:

    • ํ•™์Šต ์™„๋ฃŒ ํ›„ ๋ชจ๋ธ์„ Google Drive์— ์ €์žฅ.

์„ฑ๋Šฅ ํ‰๊ฐ€ ๋ฐ ํ…Œ์ŠคํŠธ

  • ํ‰๊ฐ€ ์ง€ํ‘œ: Accuracy, F1 score (macro, weighted) ๊ณ„์‚ฐ.
  • ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ ํ‰๊ฐ€: ํ•™์Šต๋œ ๋ชจ๋ธ์„ ์ด์šฉํ•ด ํ…Œ์ŠคํŠธ ๋ฐ์ดํ„ฐ์…‹ ํ‰๊ฐ€.

2. ๋ชจ๋ธ ์‚ฌ์šฉ ๋ฐฉ๋ฒ•

์‚ฌ์ „ ์ค€๋น„

  • HuggingFace Hub์—์„œ ํ•™์Šต๋œ ๋ชจ๋ธ์„ ๋ถˆ๋Ÿฌ์™€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
  • ๋ชจ๋ธ ๋ฐ ํ† ํฌ๋‚˜์ด์ €๋Š” monologg/kobert ๊ธฐ๋ฐ˜์ด๋ฉฐ, ๋ถ„๋ฅ˜ ๋ ˆ์ด๋ธ”์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:
    • Anger: ๐Ÿ˜ก
    • Fear: ๐Ÿ˜จ
    • Happy: ๐Ÿ˜Š
    • Tender: ๐Ÿฅฐ
    • Sad: ๐Ÿ˜ข

์‚ฌ์šฉ ์˜ˆ์‹œ

  1. ๋‹จ์ˆœ ๋ฌธ์žฅ ์ž…๋ ฅ ๊ฐ์ • ๋ถ„์„:

    • ์‚ฌ์šฉ์ž๊ฐ€ ์ž…๋ ฅํ•œ ํ…์ŠคํŠธ์— ๋Œ€ํ•ด ๋ชจ๋ธ์ด ๊ฐ์ •์„ ์˜ˆ์ธกํ•˜๊ณ , ๊ฐ ๊ฐ์ •์˜ ํ™•๋ฅ ์„ ํ•จ๊ป˜ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค.
  2. ์—‘์…€ ํŒŒ์ผ์—์„œ ๊ฐ์ • ๋ถ„์„:

    • ์—‘์…€ ํŒŒ์ผ์—์„œ ์ง€์ •ํ•œ ํ…์ŠคํŠธ ์—ด๊ณผ ํ–‰ ๋ฒ”์œ„๋ฅผ ์ฝ์–ด์™€, ํ•ด๋‹น ํ…์ŠคํŠธ๋“ค์— ๋Œ€ํ•ด ๊ฐ์ •์„ ๋ถ„๋ฅ˜ํ•˜๊ณ  ๊ฒฐ๊ณผ๋ฅผ ์ถœ๋ ฅํ•ฉ๋‹ˆ๋‹ค.

์ฝ”๋“œ ์‚ฌ์šฉ ์˜ˆ์‹œ

# ํ† ํฌ๋‚˜์ด์ € ๋ฐ ๋ชจ๋ธ ๋กœ๋“œ
from transformers import AutoTokenizer, AutoModelForSequenceClassification

# KoBERT ํ† ํฌ๋‚˜์ด์ €์™€ ๋ชจ๋ธ ๋กœ๋“œ
tokenizer = AutoTokenizer.from_pretrained("monologg/kobert", trust_remote_code=True)
model = AutoModelForSequenceClassification.from_pretrained("rkdaldus/ko-sent5-classification")

# ์‚ฌ์šฉ์ž ์ž…๋ ฅ ํ…์ŠคํŠธ ๊ฐ์ • ๋ถ„์„
text = "์˜ค๋Š˜ ์ •๋ง ํ–‰๋ณตํ•ด!"
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
with torch.no_grad():
    outputs = model(**inputs)
predicted_label = torch.argmax(outputs.logits, dim=1).item()

# ๊ฐ์ • ๋ ˆ์ด๋ธ” ์ •์˜
emotion_labels = {
    0: ("Angry", "๐Ÿ˜ก"),
    1: ("Fear", "๐Ÿ˜จ"),
    2: ("Happy", "๐Ÿ˜Š"),
    3: ("Tender", "๐Ÿฅฐ"),
    4: ("Sad", "๐Ÿ˜ข")
}

# ์˜ˆ์ธก๋œ ๊ฐ์ • ์ถœ๋ ฅ
print(f"์˜ˆ์ธก๋œ ๊ฐ์ •: {emotion_labels[predicted_label][0]} {emotion_labels[predicted_label][1]}")
Downloads last month
105
Safetensors
Model size
92.2M params
Tensor type
F32
ยท
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.

Model tree for rkdaldus/ko-sent5-classification

Base model

monologg/kobert
Finetuned
(7)
this model