File size: 12,394 Bytes
7553492
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
# Korean Claim Detection Model for Fact-Checking

## ๋ชจ๋ธ ์†Œ๊ฐœ (Model Description)

์ด ๋ชจ๋ธ์€ **ํ•œ๊ตญ์–ด ๋ฌธ์žฅ์—์„œ ํŒฉํŠธ์ฒดํฌ๊ฐ€ ํ•„์š”ํ•œ ์ฃผ์žฅ(claim)์„ ์ž๋™์œผ๋กœ ํƒ์ง€**ํ•˜๋Š” ์ด์ง„ ๋ถ„๋ฅ˜ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

๋‰ด์Šค ๊ธฐ์‚ฌ, ์ •์น˜ ํ† ๋ก , ์†Œ์…œ ๋ฏธ๋””์–ด ๊ฒŒ์‹œ๋ฌผ ๋“ฑ์—์„œ ๊ฒ€์ฆ์ด ํ•„์š”ํ•œ ์ฃผ์žฅ์„ ์‹๋ณ„ํ•˜์—ฌ, ํŒฉํŠธ์ฒดํ‚น ์›Œํฌํ”Œ๋กœ์šฐ์˜ ์ฒซ ๋‹จ๊ณ„๋ฅผ ์ž๋™ํ™”ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

This model **automatically detects claims that require fact-checking** in Korean sentences. It can identify verifiable claims in news articles, political debates, and social media posts, automating the first step of the fact-checking workflow.

- **Base Model**: [beomi/KcELECTRA-base-v2022](https://huggingface.co/beomi/KcELECTRA-base-v2022)
- **Task**: Claim Detection (Checkworthy Sentence Classification)
- **Language**: Korean (ํ•œ๊ตญ์–ด)
- **Labels**:
  - `0`: ํŒฉํŠธ์ฒดํฌ๊ฐ€ ๋ถˆํ•„์š”ํ•œ ๋ฌธ์žฅ (Non-checkworthy)
  - `1`: ํŒฉํŠธ์ฒดํฌ๊ฐ€ ํ•„์š”ํ•œ ์ฃผ์žฅ (Checkworthy claim)

## ๋ชจ๋ธ ๋ชฉํ‘œ (Model Objective)

์ž…๋ ฅ๋œ ํ•œ๊ตญ์–ด ๋ฌธ์žฅ์„ ๋ถ„์„ํ•˜์—ฌ ๋‹ค์Œ์„ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค:
- **๊ฒ€์ฆ ๊ฐ€๋Šฅํ•œ ์‚ฌ์‹ค์  ์ฃผ์žฅ**์ธ์ง€
- **ํŒฉํŠธ์ฒดํ‚น์ด ํ•„์š”ํ•œ ์ •๋„**๋Š” ์–ผ๋งˆ๋‚˜ ๋˜๋Š”์ง€

This model analyzes Korean sentences to determine:
- Whether they contain **verifiable factual claims**
- The **degree to which fact-checking is needed**

### ํŒฉํŠธ์ฒดํฌ๊ฐ€ ํ•„์š”ํ•œ ์ฃผ์žฅ์˜ ์˜ˆ์‹œ (Checkworthy Claim Examples)

โœ… **Label 1 (Checkworthy)**:
- "์ฒญ๋…„ ์‹ค์—…๋ฅ ์ด ์ง€๋‚œ 3๋…„๊ฐ„ ๊ณ„์† ์ƒ์Šนํ–ˆ์Šต๋‹ˆ๋‹ค"
- "์šฐ๋ฆฌ๋‚˜๋ผ GDP ์„ฑ์žฅ๋ฅ ์€ OECD ํ‰๊ท ์„ ๋„˜์–ด์„ฐ์Šต๋‹ˆ๋‹ค"
- "์ด ์ •์ฑ…์œผ๋กœ ์ผ์ž๋ฆฌ๊ฐ€ 100๋งŒ ๊ฐœ ์ฐฝ์ถœ๋  ๊ฒƒ์ž…๋‹ˆ๋‹ค"

โŒ **Label 0 (Non-checkworthy)**:
- "์˜ค๋Š˜ ํ† ๋ก ํšŒ๋Š” SBS ์ƒ์•”๋™ ์ŠคํŠœ๋””์˜ค์—์„œ ์ง„ํ–‰ํ•˜๊ณ  ์žˆ๊ณ ์š”"
- "๊ตญ๋ฏผ ์—ฌ๋Ÿฌ๋ถ„๊ป˜ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค"
- "์ œ ์ƒ๊ฐ์—๋Š” ์ด ์ •์ฑ…์ด ์ข‹์€ ๊ฒƒ ๊ฐ™์Šต๋‹ˆ๋‹ค"

## ๋ฐ์ดํ„ฐ์…‹ (Dataset)

### ๋ฐ์ดํ„ฐ ์ถœ์ฒ˜
- **Source**: CLEF CheckThat! Lab 2024
- **Task**: Task 1 - Check-Worthiness Estimation
- **Original Dataset**: English political debates and speeches
- **Translation**: Machine-translated to Korean for training

### ๋ฐ์ดํ„ฐ์…‹ ํฌ๊ธฐ
- **Training Set**: 22,501 samples
- **Validation Set**: 1,032 samples
- **Test Set**: 318 samples

### ๋ฐ์ดํ„ฐ ํŠน์„ฑ
- ์ •์น˜ ํ† ๋ก , ์—ฐ์„ค๋ฌธ, ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ์ถ”์ถœ๋œ ๋ฌธ์žฅ
- ์ „๋ฌธ ํŒฉํŠธ์ฒด์ปค๋“ค์ด ๋ ˆ์ด๋ธ”๋งํ•œ ๊ณ ํ’ˆ์งˆ ๋ฐ์ดํ„ฐ
- ํด๋ž˜์Šค ๋ถˆ๊ท ํ˜•: Label 0 (65%) vs Label 1 (35%)

## ํ•™์Šต ์„ธ๋ถ€์‚ฌํ•ญ (Training Details)

### ํ•™์Šต ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ
- **Epochs**: 5
- **Batch Size (Train)**: 32
- **Batch Size (Eval)**: 64
- **Learning Rate**: 3e-05
- **Weight Decay**: 0.01
- **Warmup Ratio**: 0.1
- **Precision**: BF16
- **Optimizer**: adamw_torch_fused
- **Max Sequence Length**: 128 tokens
- **Seed**: 42

### ํ•™์Šต ํ™˜๊ฒฝ
- **GPU**: NVIDIA GeForce RTX 4090 (24GB)
- **Training Time**: 1.87 minutes
- **Framework**: Hugging Face Transformers
- **Early Stopping**: Patience 3 (based on F1 score)

## ์„ฑ๋Šฅ (Performance)

### Validation Metrics
- **Accuracy**: 97.58%
- **F1 Score**: 94.80%
- **Precision**: 93.83%
- **Recall**: 95.80%

### Test Metrics
- **Accuracy**: 89.31%
- **F1 Score**: 82.65%
- **Precision**: 92.05%
- **Recall**: 75.00%

### Confusion Matrix (Test Set)
```
           Predicted
           0      1
Actual 0   203    7    (96.7% ์ •ํ™•๋„)
       1    27    81   (75.0% ์žฌํ˜„์œจ)
```

**์„ฑ๋Šฅ ํ•ด์„**:
- **๋†’์€ Precision (92.05%)**: ๋ชจ๋ธ์ด "checkworthy"๋ผ๊ณ  ์˜ˆ์ธกํ•œ ๋ฌธ์žฅ์˜ 92%๊ฐ€ ์‹ค์ œ๋กœ ํŒฉํŠธ์ฒดํฌ๊ฐ€ ํ•„์š”
- **์ ์ ˆํ•œ Recall (75.00%)**: ์‹ค์ œ checkworthy ๋ฌธ์žฅ์˜ 75%๋ฅผ ํƒ์ง€
- **๋‚ฎ์€ False Positive (7๊ฐœ)**: ๋ถˆํ•„์š”ํ•œ ํŒฉํŠธ์ฒดํฌ ์š”์ฒญ ์ตœ์†Œํ™”

## ์‚ฌ์šฉ ๋ฐฉ๋ฒ• (How to Use)

### 1. ์„ค์น˜ (Installation)

```bash
pip install transformers torch
```

### 2. ๋ชจ๋ธ ๋กœ๋“œ (Loading the Model)

```python
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# ๋ชจ๋ธ ๋กœ๋“œ
model_name = "jonghhhh/claim_factcheck"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSequenceClassification.from_pretrained(model_name)

# GPU ์‚ฌ์šฉ (์„ ํƒ์‚ฌํ•ญ)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

print(f"โœ… ๋ชจ๋ธ ๋กœ๋“œ ์™„๋ฃŒ! (Device: {device})")
```

### 3. ์ถ”๋ก  ์˜ˆ์‹œ (Inference Example)

#### ๋‹จ์ผ ๋ฌธ์žฅ ๋ถ„๋ฅ˜

```python
def predict_claim(text):
    """
    ์ž…๋ ฅ ๋ฌธ์žฅ์ด ํŒฉํŠธ์ฒดํฌ๊ฐ€ ํ•„์š”ํ•œ ์ฃผ์žฅ์ธ์ง€ ํŒ๋‹จํ•ฉ๋‹ˆ๋‹ค.

    Args:
        text (str): ๋ถ„์„ํ•  ํ•œ๊ตญ์–ด ๋ฌธ์žฅ

    Returns:
        dict: {
            'text': ์ž…๋ ฅ ๋ฌธ์žฅ,
            'is_checkworthy': True/False,
            'confidence': 0.0~1.0 (ํ™•์‹ ๋„),
            'label': 0 ๋˜๋Š” 1,
            'probabilities': {'non_checkworthy': 0.xx, 'checkworthy': 0.xx}
        }
    """
    # ํ† ํฌ๋‚˜์ด์ง•
    inputs = tokenizer(
        text,
        truncation=True,
        max_length=128,
        return_tensors="pt"
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}

    # ์ถ”๋ก 
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=-1)
        predicted_label = torch.argmax(probs, dim=-1).item()
        confidence = probs[0][predicted_label].item()

    return {
        'text': text,
        'is_checkworthy': bool(predicted_label),
        'confidence': confidence,
        'label': predicted_label,
        'probabilities': {
            'non_checkworthy': probs[0][0].item(),
            'checkworthy': probs[0][1].item()
        }
    }

# ์‚ฌ์šฉ ์˜ˆ์‹œ
examples = [
    "์˜ค๋Š˜ ํ† ๋ก ํšŒ๋Š” SBS ์ƒ์•”๋™ ์ŠคํŠœ๋””์˜ค์—์„œ ์ง„ํ–‰ํ•˜๊ณ  ์žˆ๊ณ ์š”.",
    "์ฒญ๋…„ ์‹ค์—…๋ฅ ์ด ์ตœ๊ทผ 3๋…„๊ฐ„ ๊ณ„์† ์ƒ์Šนํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.",
    "์šฐ๋ฆฌ๋‚˜๋ผ GDP ์„ฑ์žฅ๋ฅ ์€ OECD ํ‰๊ท ์„ ๋„˜์–ด์„ฐ์Šต๋‹ˆ๋‹ค.",
    "๊ตญ๋ฏผ ์—ฌ๋Ÿฌ๋ถ„๊ป˜ ์ง„์‹ฌ์œผ๋กœ ๊ฐ์‚ฌ๋“œ๋ฆฝ๋‹ˆ๋‹ค."
]

for text in examples:
    result = predict_claim(text)
    print(f"\n๐Ÿ“ ์ž…๋ ฅ: {result['text']}")
    print(f"{'๐Ÿ” ํŒฉํŠธ์ฒดํฌ ํ•„์š”' if result['is_checkworthy'] else 'โœ… ํŒฉํŠธ์ฒดํฌ ๋ถˆํ•„์š”'}")
    print(f"ํ™•์‹ ๋„: {result['confidence']*100:.1f}%")
    print(f"์ƒ์„ธ ํ™•๋ฅ : Non-CW {result['probabilities']['non_checkworthy']*100:.1f}% | CW {result['probabilities']['checkworthy']*100:.1f}%")
```

**์ถœ๋ ฅ ์˜ˆ์‹œ**:
```
๐Ÿ“ ์ž…๋ ฅ: ์ฒญ๋…„ ์‹ค์—…๋ฅ ์ด ์ตœ๊ทผ 3๋…„๊ฐ„ ๊ณ„์† ์ƒ์Šนํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿ” ํŒฉํŠธ์ฒดํฌ ํ•„์š”
ํ™•์‹ ๋„: 94.3%
์ƒ์„ธ ํ™•๋ฅ : Non-CW 5.7% | CW 94.3%

๐Ÿ“ ์ž…๋ ฅ: ์˜ค๋Š˜ ํ† ๋ก ํšŒ๋Š” SBS ์ƒ์•”๋™ ์ŠคํŠœ๋””์˜ค์—์„œ ์ง„ํ–‰ํ•˜๊ณ  ์žˆ๊ณ ์š”.
โœ… ํŒฉํŠธ์ฒดํฌ ๋ถˆํ•„์š”
ํ™•์‹ ๋„: 98.2%
์ƒ์„ธ ํ™•๋ฅ : Non-CW 98.2% | CW 1.8%
```

#### ๋ฐฐ์น˜ ์ฒ˜๋ฆฌ (Batch Processing)

```python
def predict_claims_batch(texts, batch_size=32):
    """
    ์—ฌ๋Ÿฌ ๋ฌธ์žฅ์„ ๋ฐฐ์น˜๋กœ ์ฒ˜๋ฆฌํ•ฉ๋‹ˆ๋‹ค.

    Args:
        texts (list): ๋ฌธ์žฅ ๋ฆฌ์ŠคํŠธ
        batch_size (int): ๋ฐฐ์น˜ ํฌ๊ธฐ

    Returns:
        list: ๊ฐ ๋ฌธ์žฅ์˜ ์˜ˆ์ธก ๊ฒฐ๊ณผ ๋ฆฌ์ŠคํŠธ
    """
    results = []

    for i in range(0, len(texts), batch_size):
        batch_texts = texts[i:i+batch_size]

        # ๋ฐฐ์น˜ ํ† ํฌ๋‚˜์ด์ง•
        inputs = tokenizer(
            batch_texts,
            truncation=True,
            max_length=128,
            padding=True,
            return_tensors="pt"
        )
        inputs = {k: v.to(device) for k, v in inputs.items()}

        # ๋ฐฐ์น˜ ์ถ”๋ก 
        with torch.no_grad():
            outputs = model(**inputs)
            probs = torch.softmax(outputs.logits, dim=-1)
            predicted_labels = torch.argmax(probs, dim=-1).cpu().numpy()

        # ๊ฒฐ๊ณผ ์ €์žฅ
        for j, text in enumerate(batch_texts):
            results.append({
                'text': text,
                'is_checkworthy': bool(predicted_labels[j]),
                'confidence': probs[j][predicted_labels[j]].item(),
                'label': int(predicted_labels[j])
            })

    return results

# ๋ฐฐ์น˜ ์ถ”๋ก  ์˜ˆ์‹œ
texts = [
    "๊ตญํšŒ์˜์› ์ •์›์„ 300๋ช…์œผ๋กœ ํ™•๋Œ€ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค.",
    "๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค.",
    "2024๋…„ ๊ฒฝ์ œ์„ฑ์žฅ๋ฅ ์ด 2.1%๋ฅผ ๊ธฐ๋กํ–ˆ์Šต๋‹ˆ๋‹ค.",
    # ... ๋” ๋งŽ์€ ๋ฌธ์žฅ๋“ค
]

batch_results = predict_claims_batch(texts)
checkworthy_claims = [r for r in batch_results if r['is_checkworthy']]
print(f"โœ… ์ด {len(texts)}๊ฐœ ๋ฌธ์žฅ ์ค‘ {len(checkworthy_claims)}๊ฐœ๊ฐ€ ํŒฉํŠธ์ฒดํฌ ํ•„์š”")
```

### 4. ์‹ค์ „ ํ™œ์šฉ ์˜ˆ์‹œ (Real-world Use Case)

```python
# ๋‰ด์Šค ๊ธฐ์‚ฌ์—์„œ ํŒฉํŠธ์ฒดํฌ ๋Œ€์ƒ ์ถ”์ถœ
def extract_checkworthy_claims(article_text, threshold=0.7):
    """
    ๊ธฐ์‚ฌ์—์„œ ํŒฉํŠธ์ฒดํฌ๊ฐ€ ํ•„์š”ํ•œ ๋ฌธ์žฅ๋“ค์„ ์ถ”์ถœํ•ฉ๋‹ˆ๋‹ค.

    Args:
        article_text (str): ๋‰ด์Šค ๊ธฐ์‚ฌ ์ „๋ฌธ
        threshold (float): checkworthy ํŒ๋‹จ ์ž„๊ณ„๊ฐ’ (0.0~1.0)

    Returns:
        list: ํŒฉํŠธ์ฒดํฌ ๋Œ€์ƒ ๋ฌธ์žฅ๋“ค
    """
    # ๋ฌธ์žฅ ๋ถ„๋ฆฌ (๊ฐ„๋‹จํ•œ ์˜ˆ์‹œ)
    sentences = [s.strip() for s in article_text.split('.') if s.strip()]

    # ๋ฐฐ์น˜ ์˜ˆ์ธก
    results = predict_claims_batch(sentences)

    # ์ž„๊ณ„๊ฐ’ ์ด์ƒ์˜ checkworthy ๋ฌธ์žฅ๋งŒ ํ•„ํ„ฐ๋ง
    checkworthy_claims = [
        r for r in results
        if r['is_checkworthy'] and r['confidence'] >= threshold
    ]

    # ํ™•์‹ ๋„ ์ˆœ์œผ๋กœ ์ •๋ ฌ
    checkworthy_claims.sort(key=lambda x: x['confidence'], reverse=True)

    return checkworthy_claims

# ์‚ฌ์šฉ ์˜ˆ์‹œ
article = """
์ •๋ถ€๋Š” ์˜ค๋Š˜ ๊ฒฝ์ œ์ •์ฑ… ๋ฐฉํ–ฅ์„ ๋ฐœํ‘œํ–ˆ์Šต๋‹ˆ๋‹ค.
์ฒญ๋…„ ์‹ค์—…๋ฅ ์ด ์ง€๋‚œํ•ด ๋Œ€๋น„ 2.3%p ๊ฐ์†Œํ–ˆ๋‹ค๊ณ  ๋ฐํ˜”์Šต๋‹ˆ๋‹ค.
์ด๋Š” ์—ญ๋Œ€ ์ตœ๋Œ€ ํญ์˜ ํ•˜๋ฝ์ž…๋‹ˆ๋‹ค.
์•ž์œผ๋กœ๋„ ์ผ์ž๋ฆฌ ์ฐฝ์ถœ์— ํž˜์“ฐ๊ฒ ๋‹ค๊ณ  ๊ฐ•์กฐํ–ˆ์Šต๋‹ˆ๋‹ค.
"""

claims = extract_checkworthy_claims(article, threshold=0.8)
print(f"๐Ÿ” ๋ฐœ๊ฒฌ๋œ ํŒฉํŠธ์ฒดํฌ ๋Œ€์ƒ: {len(claims)}๊ฐœ\n")

for i, claim in enumerate(claims, 1):
    print(f"{i}. {claim['text']}")
    print(f"   ํ™•์‹ ๋„: {claim['confidence']*100:.1f}%\n")
```

## ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜ (Model Architecture)

- **Model Type**: ELECTRA (Efficiently Learning an Encoder that Classifies Token Replacements Accurately)
- **Hidden Size**: 768
- **Number of Layers**: 12
- **Number of Attention Heads**: 12
- **Vocabulary Size**: 32,000
- **Max Sequence Length**: 128 tokens
- **Classification Head**: Linear layer (768 โ†’ 2)

## ํ•œ๊ณ„ ๋ฐ ๊ณ ๋ ค์‚ฌํ•ญ (Limitations)

1. **๋„๋ฉ”์ธ ํŠนํ™”**: ์ •์น˜/๋‰ด์Šค ๋„๋ฉ”์ธ์— ์ตœ์ ํ™”๋˜์–ด ์žˆ์–ด, ์ผ์ƒ ๋Œ€ํ™”๋‚˜ ๊ธฐ์ˆ  ๋ฌธ์„œ์—๋Š” ์„ฑ๋Šฅ์ด ๋–จ์–ด์งˆ ์ˆ˜ ์žˆ์Œ
2. **๊ธธ์ด ์ œํ•œ**: ์ตœ๋Œ€ 128 ํ† ํฐ๊นŒ์ง€๋งŒ ์ฒ˜๋ฆฌ ๊ฐ€๋Šฅ (์•ฝ 100-150 ๋‹จ์–ด)
3. **๊ธฐ๊ณ„ ๋ฒˆ์—ญ ๋ฐ์ดํ„ฐ**: ์˜์–ด์—์„œ ๋ฒˆ์—ญ๋œ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šต๋˜์–ด ์ž์—ฐ์Šค๋Ÿฌ์šด ํ•œ๊ตญ์–ด ํ‘œํ˜„์—์„œ ์„ฑ๋Šฅ ์ฐจ์ด ๊ฐ€๋Šฅ
4. **์ด์ง„ ๋ถ„๋ฅ˜**: Checkworthy ์ •๋„๋ฅผ 0/1๋กœ๋งŒ ๋ถ„๋ฅ˜ (์„ธ๋ฐ€ํ•œ ์ ์ˆ˜ ์ œ๊ณต ์•ˆ ํ•จ)
5. **False Negative**: ์‹ค์ œ ์ฃผ์žฅ์˜ 25%๋ฅผ ๋†“์น  ์ˆ˜ ์žˆ์Œ (Recall 75%)

## ๊ฐœ์„  ๋ฐฉํ–ฅ (Future Improvements)

- [ ] ํ•œ๊ตญ์–ด ๋„ค์ดํ‹ฐ๋ธŒ ํŒฉํŠธ์ฒดํฌ ๋ฐ์ดํ„ฐ์…‹์œผ๋กœ ์ถ”๊ฐ€ ํ•™์Šต
- [ ] ๋” ๊ธด ๋ฌธ๋งฅ ์ฒ˜๋ฆฌ๋ฅผ ์œ„ํ•œ ๋ชจ๋ธ ์—…๊ทธ๋ ˆ์ด๋“œ (max_length 256+)
- [ ] ๋‹ค์ค‘ ํด๋ž˜์Šค ๋ถ„๋ฅ˜ (checkworthy ์ ์ˆ˜๋ฅผ 0-5 ์ฒ™๋„๋กœ)
- [ ] ์ฃผ์žฅ์˜ ์ฃผ์ œ ์นดํ…Œ๊ณ ๋ฆฌ ๋ถ„๋ฅ˜ ๊ธฐ๋Šฅ ์ถ”๊ฐ€

## ๋ผ์ด์„ ์Šค (License)

์ด ๋ชจ๋ธ์€ ๋ฒ ์ด์Šค ๋ชจ๋ธ์ธ [beomi/KcELECTRA-base-v2022](https://huggingface.co/beomi/KcELECTRA-base-v2022)์˜ ๋ผ์ด์„ ์Šค๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค.

## ์ธ์šฉ (Citation)

์ด ๋ชจ๋ธ์„ ์—ฐ๊ตฌ๋‚˜ ํ”„๋กœ์ ํŠธ์— ์‚ฌ์šฉํ•˜์‹ ๋‹ค๋ฉด ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ธ์šฉํ•ด์ฃผ์„ธ์š”:

```bibtex
@misc{korean-claim-factcheck-2025,
  author = {Jonghhhh},
  title = {Korean Claim Detection Model for Fact-Checking},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/jonghhhh/claim_factcheck}},
  note = {Based on KcELECTRA-base-v2022}
}
```

## ์ฐธ๊ณ  ์ž๋ฃŒ (References)

- **Base Model**: [beomi/KcELECTRA-base-v2022](https://huggingface.co/beomi/KcELECTRA-base-v2022)
- **Dataset**: [CLEF CheckThat! Lab 2024](https://clef2025.clef-initiative.eu/index.php?page=Pages/Labs/CheckThat.html)
- **Paper**: [CheckThat! Lab: Check-Worthiness, Subjectivity, and Persuasion](https://link.springer.com/chapter/10.1007/978-3-031-13643-6_24)

## ์—ฐ๋ฝ์ฒ˜ (Contact)

์งˆ๋ฌธ์ด๋‚˜ ํ”ผ๋“œ๋ฐฑ์ด ์žˆ์œผ์‹œ๋ฉด Issues๋ฅผ ํ†ตํ•ด ๋‚จ๊ฒจ์ฃผ์„ธ์š”!

---

**Tags**: `claim-detection`, `fact-checking`, `korean`, `electra`, `text-classification`, `checkworthy`, `misinformation-detection`