File size: 2,404 Bytes
3c6b13d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
---
language: ko
datasets:
- mc4
license: apache-2.0
---

# ByT5-Korean - large

ByT5-Korean is a Korean specific extension of Google's [ByT5](https://github.com/google-research/byt5).

A Korean syllable has three components (called Jamo): a beginning consonant, a middle vowel, and an optional final consonant; they are like individual characters of alphabet.
While the ByT5's utf-8 encoding allows generic encoding for multiple languages, it is unnatural for Korean because it splits the bits representation of each Jamo in the middle.

ByT5-Korean extends ByT5's utf-8 encoding with special care for Korean syllables; each Jamo is represented with a extra token.
ByT5-Korean was pre-trained on [mC4](https://www.tensorflow.org/datasets/catalog/c4#c4multilingual) with 70% Korean and 30% English.

## Encoding Scheme
```text
id: token
0: <pad>
1: <eos>
2: <unk>
3~258: utf-8 encoding
259~277: beginning consonants(μ΄ˆμ„±), 19개(γ„±γ„²γ„΄γ„·γ„Έγ„Ήγ…γ…‚γ…ƒγ……γ…†γ…‡γ…ˆγ…‰γ…Šγ…‹γ…Œγ…γ…Ž)
278~298: middle vowel(쀑성), 21개(γ…γ…γ…‘γ…’γ…“γ…”γ…•γ…–γ…—γ…˜γ…™γ…šγ…›γ…œγ…γ…žγ…Ÿγ… γ…‘γ…’γ…£)
299~326: final consonant(μ’…μ„±), 무쒅성+27개(γ„±γ„²γ„³γ„΄γ„΅γ„Άγ„·γ„Ήγ„Ίγ„»γ„Όγ„½γ„Ύγ„Ώγ…€γ…γ…‚γ…„γ……γ…†γ…‡γ…ˆγ…Šγ…‹γ…Œγ…γ…Ž)
327~384: from <extra_id_0> to <extra_id_57>
```

## Example Inference

```python
import torch
from tokenizer import ByT5KoreanTokenizer # https://huggingface.co/everdoubling/byt5-Korean-large/blob/main/tokenizer.py
from transformers import T5ForConditionalGeneration

tokenizer_jamo = ByT5KoreanTokenizer()
model = T5ForConditionalGeneration.from_pretrained('everdoubling/byt5-Korean-large')

input_sentence = 'ν•œκ΅­μ–΄ μœ„ν‚€λ°±κ³Ό(μ˜μ–΄: Korean Wikipedia)λŠ” ν•œκ΅­μ–΄λ‘œ μš΄μ˜λ˜λŠ” μœ„ν‚€λ°±κ³Όμ˜ λ‹€μ–Έμ–΄νŒ κ°€μš΄λ° ν•˜λ‚˜λ‘œμ„œ, 2002λ…„ 10μ›” 11일에 <extra_id_0>. λ˜ν•œ ν˜„μž¬ ν•œκ΅­μ–΄ μœ„ν‚€λ°±κ³Όμ—λŠ” λ„˜κ²¨μ£ΌκΈ°, ν† λ‘ , κ·Έλ¦Ό λ“± νŽ˜μ΄μ§€λ‘œ λΆˆλ¦¬λŠ” λͺ¨λ“  λ¬Έμ„œλ₯Ό ν¬ν•¨ν•˜λ©΄ 총 2,629,860κ°œκ°€ <extra_id_1>λ˜μ–΄ 있으며, λ„˜κ²¨μ£ΌκΈ°λ₯Ό ν¬ν•¨ν•œ 일반 λ¬Έμ„œ μˆ˜λŠ” 1,278,560개,[1] 그쀑 λ„˜κ²¨μ£ΌκΈ°, 막닀λ₯Έ λ¬Έμ„œλ₯Ό μ œμ™Έν•œ 일반 λ¬Έμ„œ μˆ˜λŠ” 573,149κ°œμ΄λ‹€.'

input_ids_jamo = tokenizer_jamo(input_sentence).input_ids
outputs_jamo = model_jamo.generate(torch.tensor([input_ids_jamo]))
print(tokenizer_jamo.decode(outputs_jamo[0]))
# <pad><extra_id_0>μ„€λ¦½λ˜μ—ˆλ‹€<extra_id_1>Δ‘Δ›
```

Additional information coming soon...