File size: 12,410 Bytes
5b4ca19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4cf9eb1
5b4ca19
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language: 
  - ko
  - en
tags:
- electra
- korean
license: "mit"
---


# KcELECTRA: Korean comments ELECTRA

** Updates on 2022.10.08 **

- KcELECTRA-base-v2022 (๊ตฌ v2022-dev) ๋ชจ๋ธ ์ด๋ฆ„์ด ๋ณ€๊ฒฝ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. --> KcELECTRA-base ๋ ˆํฌ์˜ `v2022`๋กœ ํ†ตํ•ฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
- ์œ„ ๋ชจ๋ธ์˜ ์„ธ๋ถ€ ์Šค์ฝ”์–ด๋ฅผ ์ถ”๊ฐ€ํ•˜์˜€์Šต๋‹ˆ๋‹ค.
- ๊ธฐ์กด KcELECTRA-base(v2021) ๋Œ€๋น„ ๋Œ€๋ถ€๋ถ„์˜ downstream task์—์„œ ~1%p ์ˆ˜์ค€์˜ ์„ฑ๋Šฅ ํ–ฅ์ƒ์ด ์žˆ์Šต๋‹ˆ๋‹ค.

---

๊ณต๊ฐœ๋œ ํ•œ๊ตญ์–ด Transformer ๊ณ„์—ด ๋ชจ๋ธ๋“ค์€ ๋Œ€๋ถ€๋ถ„ ํ•œ๊ตญ์–ด ์œ„ํ‚ค, ๋‰ด์Šค ๊ธฐ์‚ฌ, ์ฑ… ๋“ฑ ์ž˜ ์ •์ œ๋œ ๋ฐ์ดํ„ฐ๋ฅผ ๊ธฐ๋ฐ˜์œผ๋กœ ํ•™์Šตํ•œ ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค. ํ•œํŽธ, ์‹ค์ œ๋กœ NSMC์™€ ๊ฐ™์€ User-Generated Noisy text domain ๋ฐ์ดํ„ฐ์…‹์€ ์ •์ œ๋˜์ง€ ์•Š์•˜๊ณ  ๊ตฌ์–ด์ฒด ํŠน์ง•์— ์‹ ์กฐ์–ด๊ฐ€ ๋งŽ์œผ๋ฉฐ, ์˜คํƒˆ์ž ๋“ฑ ๊ณต์‹์ ์ธ ๊ธ€์“ฐ๊ธฐ์—์„œ ๋‚˜ํƒ€๋‚˜์ง€ ์•Š๋Š” ํ‘œํ˜„๋“ค์ด ๋นˆ๋ฒˆํ•˜๊ฒŒ ๋“ฑ์žฅํ•ฉ๋‹ˆ๋‹ค.

KcELECTRA๋Š” ์œ„์™€ ๊ฐ™์€ ํŠน์„ฑ์˜ ๋ฐ์ดํ„ฐ์…‹์— ์ ์šฉํ•˜๊ธฐ ์œ„ํ•ด, ๋„ค์ด๋ฒ„ ๋‰ด์Šค์—์„œ ๋Œ“๊ธ€๊ณผ ๋Œ€๋Œ“๊ธ€์„ ์ˆ˜์ง‘ํ•ด, ํ† ํฌ๋‚˜์ด์ €์™€ ELECTRA๋ชจ๋ธ์„ ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šตํ•œ Pretrained ELECTRA ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.

๊ธฐ์กด KcBERT ๋Œ€๋น„ ๋ฐ์ดํ„ฐ์…‹ ์ฆ๊ฐ€ ๋ฐ vocab ํ™•์žฅ์„ ํ†ตํ•ด ์ƒ๋‹นํ•œ ์ˆ˜์ค€์œผ๋กœ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.

KcELECTRA๋Š” Huggingface์˜ Transformers ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ†ตํ•ด ๊ฐ„ํŽธํžˆ ๋ถˆ๋Ÿฌ์™€ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. (๋ณ„๋„์˜ ํŒŒ์ผ ๋‹ค์šด๋กœ๋“œ๊ฐ€ ํ•„์š”ํ•˜์ง€ ์•Š์Šต๋‹ˆ๋‹ค.)

```
๐Ÿ’ก NOTE ๐Ÿ’ก 
General Corpus๋กœ ํ•™์Šตํ•œ KoELECTRA๊ฐ€ ๋ณดํŽธ์ ์ธ task์—์„œ๋Š” ์„ฑ๋Šฅ์ด ๋” ์ž˜ ๋‚˜์˜ฌ ๊ฐ€๋Šฅ์„ฑ์ด ๋†’์Šต๋‹ˆ๋‹ค.
KcBERT/KcELECTRA๋Š” User genrated, Noisy text์— ๋Œ€ํ•ด์„œ ๋ณด๋‹ค ์ž˜ ๋™์ž‘ํ•˜๋Š” PLM์ž…๋‹ˆ๋‹ค.
```

## KcELECTRA Performance

- Finetune ์ฝ”๋“œ๋Š” https://github.com/Beomi/KcBERT-finetune ์—์„œ ์ฐพ์•„๋ณด์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
- ํ•ด๋‹น Repo์˜ ๊ฐ Checkpoint ํด๋”์—์„œ Step๋ณ„ ์„ธ๋ถ€ ์Šค์ฝ”์–ด๋ฅผ ํ™•์ธํ•˜์‹ค ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

|                    | Size<br/>(์šฉ๋Ÿ‰) | **NSMC**<br/>(acc) | **Naver NER**<br/>(F1) | **PAWS**<br/>(acc) | **KorNLI**<br/>(acc) | **KorSTS**<br/>(spearman) | **Question Pair**<br/>(acc) | **KorQuaD (Dev)**<br/>(EM/F1) |
| :----------------- | :-------------: | :----------------: | :--------------------: | :----------------: | :------------------: | :-----------------------: | :-------------------------: | :---------------------------: |
| **KcELECTRA-base-v2022** |      475M       |     **91.97**      |         87.35          |       76.50        |        82.12         |           83.67           |          95.12          |         69.00 / 90.40         |
| **KcELECTRA-base** |      475M       |     91.71      |         86.90          |       74.80        |        81.65         |           82.65           |          **95.78**          |         70.60 / 90.11         |
| KcBERT-Base        |      417M       |       89.62        |         84.34          |       66.95        |        74.85         |           75.57           |            93.93            |         60.25 / 84.39         |
| KcBERT-Large       |      1.2G       |       90.68        |         85.53          |       70.15        |        76.99         |           77.49           |            94.06            |         62.16 / 86.64         |
| KoBERT             |      351M       |       89.63        |         86.11          |       80.65        |        79.00         |           79.64           |            93.93            |         52.81 / 80.27         |
| XLM-Roberta-Base   |      1.03G      |       89.49        |         86.26          |       82.95        |        79.92         |           79.09           |            93.53            |         64.70 / 88.94         |
| HanBERT            |      614M       |       90.16        |         87.31          |       82.40        |        80.89         |           83.33           |            94.19            |         78.74 / 92.02         |
| KoELECTRA-Base     |      423M       |       90.21        |         86.87          |       81.90        |        80.85         |           83.21           |            94.20            |         61.10 / 89.59         |
| KoELECTRA-Base-v2  |      423M       |       89.70        |         87.02          |       83.90        |        80.61         |           84.30           |            94.72            |         84.34 / 92.58         |
| KoELECTRA-Base-v3  |      423M       |       90.63        |       **88.11**        |     **84.45**      |      **82.24**       |         **85.53**         |            95.25            |       **84.83 / 93.45**       |
| DistilKoBERT       |      108M       |       88.41        |         84.13          |       62.55        |        70.55         |           73.21           |            92.48            |         54.12 / 77.80         |


\*HanBERT์˜ Size๋Š” Bert Model๊ณผ Tokenizer DB๋ฅผ ํ•ฉ์นœ ๊ฒƒ์ž…๋‹ˆ๋‹ค.

\***config์˜ ์„ธํŒ…์„ ๊ทธ๋Œ€๋กœ ํ•˜์—ฌ ๋Œ๋ฆฐ ๊ฒฐ๊ณผ์ด๋ฉฐ, hyperparameter tuning์„ ์ถ”๊ฐ€์ ์œผ๋กœ ํ•  ์‹œ ๋” ์ข‹์€ ์„ฑ๋Šฅ์ด ๋‚˜์˜ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.**

## How to use

### Requirements

- `pytorch ~= 1.8.0`
- `transformers ~= 4.11.3`
- `emoji ~= 0.6.0`
- `soynlp ~= 0.0.493`

### Default usage

```python
from transformers import AutoTokenizer, AutoModel
  
tokenizer = AutoTokenizer.from_pretrained("beomi/KcELECTRA-base")
model = AutoModel.from_pretrained("beomi/KcELECTRA-base")
```

> ๐Ÿ’ก ์ด์ „ KcBERT ๊ด€๋ จ ์ฝ”๋“œ๋“ค์—์„œ `AutoTokenizer`, `AutoModel` ์„ ์‚ฌ์šฉํ•œ ๊ฒฝ์šฐ `.from_pretrained("beomi/kcbert-base")` ๋ถ€๋ถ„์„ `.from_pretrained("beomi/KcELECTRA-base")` ๋กœ๋งŒ ๋ณ€๊ฒฝํ•ด์ฃผ์‹œ๋ฉด ์ฆ‰์‹œ ์‚ฌ์šฉ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค.

### Pretrain & Finetune Colab ๋งํฌ ๋ชจ์Œ 

#### Pretrain Data

- KcBERTํ•™์Šต์— ์‚ฌ์šฉํ•œ ๋ฐ์ดํ„ฐ + ์ดํ›„ 2021.03์›” ์ดˆ๊นŒ์ง€ ์ˆ˜์ง‘ํ•œ ๋Œ“๊ธ€
  - ์•ฝ 17GB
  - ๋Œ“๊ธ€-๋Œ€๋Œ“๊ธ€์„ ๋ฌถ์€ ๊ธฐ๋ฐ˜์œผ๋กœ Document ๊ตฌ์„ฑ

#### Pretrain Code

- https://github.com/KLUE-benchmark/KLUE-ELECTRA Repo๋ฅผ ํ†ตํ•œ Pretrain

#### Finetune Code

- https://github.com/Beomi/KcBERT-finetune Repo๋ฅผ ํ†ตํ•œ Finetune ๋ฐ ์Šค์ฝ”์–ด ๋น„๊ต

#### Finetune Samples

- NSMC with PyTorch-Lightning 1.3.0, GPU, Colab <a href="https://colab.research.google.com/drive/1Hh63kIBAiBw3Hho--BvfdUWLu-ysMFF0?usp=sharing">
  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
</a>


## Train Data & Preprocessing

### Raw Data

ํ•™์Šต ๋ฐ์ดํ„ฐ๋Š” 2019.01.01 ~ 2021.03.09 ์‚ฌ์ด์— ์ž‘์„ฑ๋œ **๋Œ“๊ธ€ ๋งŽ์€ ๋‰ด์Šค/ํ˜น์€ ์ „์ฒด ๋‰ด์Šค** ๊ธฐ์‚ฌ๋“ค์˜ **๋Œ“๊ธ€๊ณผ ๋Œ€๋Œ“๊ธ€**์„ ๋ชจ๋‘ ์ˆ˜์ง‘ํ•œ ๋ฐ์ดํ„ฐ์ž…๋‹ˆ๋‹ค.

๋ฐ์ดํ„ฐ ์‚ฌ์ด์ฆˆ๋Š” ํ…์ŠคํŠธ๋งŒ ์ถ”์ถœ์‹œ **์•ฝ 17.3GB์ด๋ฉฐ, 1์–ต8์ฒœ๋งŒ๊ฐœ ์ด์ƒ์˜ ๋ฌธ์žฅ**์œผ๋กœ ์ด๋ค„์ ธ ์žˆ์Šต๋‹ˆ๋‹ค.

> KcBERT๋Š” 2019.01-2020.06์˜ ํ…์ŠคํŠธ๋กœ, ์ •์ œ ํ›„ ์•ฝ 9์ฒœ๋งŒ๊ฐœ ๋ฌธ์žฅ์œผ๋กœ ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค.

### Preprocessing

PLM ํ•™์Šต์„ ์œ„ํ•ด์„œ ์ „์ฒ˜๋ฆฌ๋ฅผ ์ง„ํ–‰ํ•œ ๊ณผ์ •์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค.

1. ํ•œ๊ธ€ ๋ฐ ์˜์–ด, ํŠน์ˆ˜๋ฌธ์ž, ๊ทธ๋ฆฌ๊ณ  ์ด๋ชจ์ง€(๐Ÿฅณ)๊นŒ์ง€!

   ์ •๊ทœํ‘œํ˜„์‹์„ ํ†ตํ•ด ํ•œ๊ธ€, ์˜์–ด, ํŠน์ˆ˜๋ฌธ์ž๋ฅผ ํฌํ•จํ•ด Emoji๊นŒ์ง€ ํ•™์Šต ๋Œ€์ƒ์— ํฌํ•จํ–ˆ์Šต๋‹ˆ๋‹ค.

   ํ•œํŽธ, ํ•œ๊ธ€ ๋ฒ”์œ„๋ฅผ `ใ„ฑ-ใ…Ž๊ฐ€-ํžฃ` ์œผ๋กœ ์ง€์ •ํ•ด `ใ„ฑ-ํžฃ` ๋‚ด์˜ ํ•œ์ž๋ฅผ ์ œ์™ธํ–ˆ์Šต๋‹ˆ๋‹ค. 

2. ๋Œ“๊ธ€ ๋‚ด ์ค‘๋ณต ๋ฌธ์ž์—ด ์ถ•์•ฝ

   `ใ…‹ใ…‹ใ…‹ใ…‹ใ…‹`์™€ ๊ฐ™์ด ์ค‘๋ณต๋œ ๊ธ€์ž๋ฅผ `ใ…‹ใ…‹`์™€ ๊ฐ™์€ ๊ฒƒ์œผ๋กœ ํ•ฉ์ณค์Šต๋‹ˆ๋‹ค.

3. Cased Model

   KcBERT๋Š” ์˜๋ฌธ์— ๋Œ€ํ•ด์„œ๋Š” ๋Œ€์†Œ๋ฌธ์ž๋ฅผ ์œ ์ง€ํ•˜๋Š” Cased model์ž…๋‹ˆ๋‹ค.

4. ๊ธ€์ž ๋‹จ์œ„ 10๊ธ€์ž ์ดํ•˜ ์ œ๊ฑฐ

   10๊ธ€์ž ๋ฏธ๋งŒ์˜ ํ…์ŠคํŠธ๋Š” ๋‹จ์ผ ๋‹จ์–ด๋กœ ์ด๋ค„์ง„ ๊ฒฝ์šฐ๊ฐ€ ๋งŽ์•„ ํ•ด๋‹น ๋ถ€๋ถ„์„ ์ œ์™ธํ–ˆ์Šต๋‹ˆ๋‹ค.

5. ์ค‘๋ณต ์ œ๊ฑฐ

   ์ค‘๋ณต์ ์œผ๋กœ ์“ฐ์ธ ๋Œ“๊ธ€์„ ์ œ๊ฑฐํ•˜๊ธฐ ์œ„ํ•ด ์™„์ „ํžˆ ์ผ์น˜ํ•˜๋Š” ์ค‘๋ณต ๋Œ“๊ธ€์„ ํ•˜๋‚˜๋กœ ํ•ฉ์ณค์Šต๋‹ˆ๋‹ค.

6. `OOO` ์ œ๊ฑฐ

   ๋„ค์ด๋ฒ„ ๋Œ“๊ธ€์˜ ๊ฒฝ์šฐ, ๋น„์†์–ด๋Š” ์ž์ฒด ํ•„ํ„ฐ๋ง์„ ํ†ตํ•ด `OOO` ๋กœ ํ‘œ์‹œํ•ฉ๋‹ˆ๋‹ค. ์ด ๋ถ€๋ถ„์„ ๊ณต๋ฐฑ์œผ๋กœ ์ œ๊ฑฐํ•˜์˜€์Šต๋‹ˆ๋‹ค.

์•„๋ž˜ ๋ช…๋ น์–ด๋กœ pip๋กœ ์„ค์น˜ํ•œ ๋’ค, ์•„๋ž˜ cleanํ•จ์ˆ˜๋กœ ํด๋ฆฌ๋‹์„ ํ•˜๋ฉด Downstream task์—์„œ ๋ณด๋‹ค ์„ฑ๋Šฅ์ด ์ข‹์•„์ง‘๋‹ˆ๋‹ค. (`[UNK]` ๊ฐ์†Œ)

```bash
pip install soynlp emoji
```

์•„๋ž˜ `clean` ํ•จ์ˆ˜๋ฅผ Text data์— ์‚ฌ์šฉํ•ด์ฃผ์„ธ์š”.

```python
import re
import emoji
from soynlp.normalizer import repeat_normalize

emojis = ''.join(emoji.UNICODE_EMOJI.keys())
pattern = re.compile(f'[^ .,?!/@$%~๏ผ…ยทโˆผ()\x00-\x7Fใ„ฑ-ใ…ฃ๊ฐ€-ํžฃ{emojis}]+')
url_pattern = re.compile(
    r'https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)')

import re
import emoji
from soynlp.normalizer import repeat_normalize

pattern = re.compile(f'[^ .,?!/@$%~๏ผ…ยทโˆผ()\x00-\x7Fใ„ฑ-ใ…ฃ๊ฐ€-ํžฃ]+')
url_pattern = re.compile(
    r'https?:\/\/(www\.)?[-a-zA-Z0-9@:%._\+~#=]{1,256}\.[a-zA-Z0-9()]{1,6}\b([-a-zA-Z0-9()@:%_\+.~#?&//=]*)')

def clean(x): 
    x = pattern.sub(' ', x)
    x = emoji.replace_emoji(x, replace='') #emoji ์‚ญ์ œ
    x = url_pattern.sub('', x)
    x = x.strip()
    x = repeat_normalize(x, num_repeats=2)
    return x
```

> ๐Ÿ’ก Finetune Score์—์„œ๋Š” ์œ„ `clean` ํ•จ์ˆ˜๋ฅผ ์ ์šฉํ•˜์ง€ ์•Š์•˜์Šต๋‹ˆ๋‹ค.

### Cleaned Data

- KcBERT ์™ธ ์ถ”๊ฐ€ ๋ฐ์ดํ„ฐ๋Š” ์ •๋ฆฌ ํ›„ ๊ณต๊ฐœ ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.


## Tokenizer, Model Train

Tokenizer๋Š” Huggingface์˜ [Tokenizers](https://github.com/huggingface/tokenizers) ๋ผ์ด๋ธŒ๋Ÿฌ๋ฆฌ๋ฅผ ํ†ตํ•ด ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค.

๊ทธ ์ค‘ `BertWordPieceTokenizer` ๋ฅผ ์ด์šฉํ•ด ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ๊ณ , Vocab Size๋Š” `30000`์œผ๋กœ ์ง„ํ–‰ํ–ˆ์Šต๋‹ˆ๋‹ค.

Tokenizer๋ฅผ ํ•™์Šตํ•˜๋Š” ๊ฒƒ์—๋Š” ์ „์ฒด ๋ฐ์ดํ„ฐ๋ฅผ ํ†ตํ•ด ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ๊ณ , ๋ชจ๋ธ์˜ General Downstream task์— ๋Œ€์‘ํ•˜๊ธฐ ์œ„ํ•ด KoELECTRA์—์„œ ์‚ฌ์šฉํ•œ Vocab์„ ๊ฒน์น˜์ง€ ์•Š๋Š” ๋ถ€๋ถ„์„ ์ถ”๊ฐ€๋กœ ๋„ฃ์–ด์ฃผ์—ˆ์Šต๋‹ˆ๋‹ค. (์‹ค์ œ๋กœ ๋‘ ๋ชจ๋ธ์ด ๊ฒน์น˜๋Š” ๋ถ€๋ถ„์€ ์•ฝ 5000ํ† ํฐ์ด์—ˆ์Šต๋‹ˆ๋‹ค.)

TPU `v3-8` ์„ ์ด์šฉํ•ด ์•ฝ 10์ผ ํ•™์Šต์„ ์ง„ํ–‰ํ–ˆ๊ณ , ํ˜„์žฌ Huggingface์— ๊ณต๊ฐœ๋œ ๋ชจ๋ธ์€ 848k step์„ ํ•™์Šตํ•œ ๋ชจ๋ธ weight๊ฐ€ ์—…๋กœ๋“œ ๋˜์–ด์žˆ์Šต๋‹ˆ๋‹ค.

(100k step๋ณ„ Checkpoint๋ฅผ ํ†ตํ•ด ์„ฑ๋Šฅ ํ‰๊ฐ€๋ฅผ ์ง„ํ–‰ํ•˜์˜€์Šต๋‹ˆ๋‹ค. ํ•ด๋‹น ๋ถ€๋ถ„์€ `KcBERT-finetune` repo๋ฅผ ์ฐธ๊ณ ํ•ด์ฃผ์„ธ์š”.)

๋ชจ๋ธ ํ•™์Šต Loss๋Š” Step์— ๋”ฐ๋ผ ์ดˆ๊ธฐ 100-200k ์‚ฌ์ด์— ๊ธ‰๊ฒฉํžˆ Loss๊ฐ€ ์ค„์–ด๋“ค๋‹ค ํ•™์Šต ์ข…๋ฃŒ๊นŒ์ง€๋„ ์ง€์†์ ์œผ๋กœ loss๊ฐ€ ๊ฐ์†Œํ•˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

![KcELECTRA-base Pretrain Loss](https://cdn.jsdelivr.net/gh/beomi/blog-img@master/2021/04/07/image-20210407201231133.png)

### KcELECTRA Pretrain Step๋ณ„ Downstream task ์„ฑ๋Šฅ ๋น„๊ต

> ๐Ÿ’ก ์•„๋ž˜ ํ‘œ๋Š” ์ „์ฒด ckpt๊ฐ€ ์•„๋‹Œ ์ผ๋ถ€์— ๋Œ€ํ•ด์„œ๋งŒ ํ…Œ์ŠคํŠธ๋ฅผ ์ง„ํ–‰ํ•œ ๊ฒฐ๊ณผ์ž…๋‹ˆ๋‹ค.

![KcELECTRA Pretrain Step๋ณ„ Downstream task ์„ฑ๋Šฅ ๋น„๊ต](https://cdn.jsdelivr.net/gh/beomi/blog-img@master/2021/04/07/image-20210407215557039.png)

- ์œ„์™€ ๊ฐ™์ด KcBERT-base, KcBERT-large ๋Œ€๋น„ **๋ชจ๋“  ๋ฐ์ดํ„ฐ์…‹์— ๋Œ€ํ•ด** KcELECTRA-base๊ฐ€ ๋” ๋†’์€ ์„ฑ๋Šฅ์„ ๋ณด์ž…๋‹ˆ๋‹ค.
- KcELECTRA pretrain์—์„œ๋„ Train step์ด ๋Š˜์–ด๊ฐ์— ๋”ฐ๋ผ ์ ์ง„์ ์œผ๋กœ ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ๋˜๋Š” ๊ฒƒ์„ ๋ณผ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

## ์ธ์šฉํ‘œ๊ธฐ/Citation

KcELECTRA๋ฅผ ์ธ์šฉํ•˜์‹ค ๋•Œ๋Š” ์•„๋ž˜ ์–‘์‹์„ ํ†ตํ•ด ์ธ์šฉํ•ด์ฃผ์„ธ์š”.

```
@misc{lee2021kcelectra,
  author = {Junbum Lee},
  title = {KcELECTRA: Korean comments ELECTRA},
  year = {2021},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/Beomi/KcELECTRA}}
}
```

๋…ผ๋ฌธ์„ ํ†ตํ•œ ์‚ฌ์šฉ ์™ธ์—๋Š” MIT ๋ผ์ด์„ผ์Šค๋ฅผ ํ‘œ๊ธฐํ•ด์ฃผ์„ธ์š”. โ˜บ๏ธ

## Acknowledgement

KcELECTRA Model์„ ํ•™์Šตํ•˜๋Š” GCP/TPU ํ™˜๊ฒฝ์€ [TFRC](https://www.tensorflow.org/tfrc?hl=ko) ํ”„๋กœ๊ทธ๋žจ์˜ ์ง€์›์„ ๋ฐ›์•˜์Šต๋‹ˆ๋‹ค.

๋ชจ๋ธ ํ•™์Šต ๊ณผ์ •์—์„œ ๋งŽ์€ ์กฐ์–ธ์„ ์ฃผ์‹  [Monologg](https://github.com/monologg/) ๋‹˜ ๊ฐ์‚ฌํ•ฉ๋‹ˆ๋‹ค :)

## Reference

### Github Repos

- [KcBERT by Beomi](https://github.com/Beomi/KcBERT)
- [BERT by Google](https://github.com/google-research/bert)
- [KoBERT by SKT](https://github.com/SKTBrain/KoBERT)
- [KoELECTRA by Monologg](https://github.com/monologg/KoELECTRA/)
- [Transformers by Huggingface](https://github.com/huggingface/transformers)
- [Tokenizers by Hugginface](https://github.com/huggingface/tokenizers)
- [ELECTRA train code by KLUE](https://github.com/KLUE-benchmark/KLUE-ELECTRA)

### Blogs

- [Monologg๋‹˜์˜ KoELECTRA ํ•™์Šต๊ธฐ](https://monologg.kr/categories/NLP/ELECTRA/)
- [Colab์—์„œ TPU๋กœ BERT ์ฒ˜์Œ๋ถ€ํ„ฐ ํ•™์Šต์‹œํ‚ค๊ธฐ - Tensorflow/Google ver.](https://beomi.github.io/2020/02/26/Train-BERT-from-scratch-on-colab-TPU-Tensorflow-ver/)