eenzeenee's picture
Update README.md
b5ac75c
---
pipeline_tag: summarization
language:
- ko
tags:
- T5
---
# t5-base-korean-summarization
This is [T5](https://huggingface.co/docs/transformers/model_doc/t5) model for korean text summarization.
- Finetuned based on ['paust/pko-t5-base'](https://huggingface.co/paust/pko-t5-base) model.
- Finetuned with 3 datasets. Specifically, it is described below.
- [Korean Paper Summarization Dataset(λ…Όλ¬Έμžλ£Œ μš”μ•½)](https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=90)
- [Korean Book Summarization Dataset(λ„μ„œμžλ£Œ μš”μ•½)](https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=93)
- [Korean Summary statement and Report Generation Dataset(μš”μ•½λ¬Έ 및 레포트 생성 데이터)](https://www.aihub.or.kr/aihubdata/data/view.do?currMenu=115&topMenu=100&aihubDataSe=realm&dataSetSn=90)
# Usage (HuggingFace Transformers)
```python
import nltk
nltk.download('punkt')
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
model = AutoModelForSeq2SeqLM.from_pretrained('eenzeenee/t5-base-korean-summarization')
tokenizer = AutoTokenizer.from_pretrained('eenzeenee/t5-base-korean-summarization')
prefix = "summarize: "
sample = """
μ•ˆλ…•ν•˜μ„Έμš”? 우리 (2ν•™λ…„)/(이 ν•™λ…„) μΉœκ΅¬λ“€ 우리 μΉœκ΅¬λ“€ 학ꡐ에 κ°€μ„œ μ§„μ§œ (2ν•™λ…„)/(이 ν•™λ…„) 이 되고 μ‹Άμ—ˆλŠ”λ° 학ꡐ에 λͺ» κ°€κ³  μžˆμ–΄μ„œ λ‹΅λ‹΅ν•˜μ£ ?
κ·Έλž˜λ„ 우리 μΉœκ΅¬λ“€μ˜ μ•ˆμ „κ³Ό 건강이 μ΅œμš°μ„ μ΄λ‹ˆκΉŒμš” μ˜€λŠ˜λΆ€ν„° μ„ μƒλ‹˜μ΄λž‘ 맀일 맀일 κ΅­μ–΄ 여행을 λ– λ‚˜λ³΄λ„λ‘ ν•΄μš”.
μ–΄/ μ‹œκ°„μ΄ 벌써 μ΄λ ‡κ²Œ λλ‚˜μš”? λŠ¦μ—ˆμ–΄μš”. λŠ¦μ—ˆμ–΄μš”. 빨리 κ΅­μ–΄ 여행을 λ– λ‚˜μ•Ό λΌμš”.
그런데 μ–΄/ ꡭ어여행을 λ– λ‚˜κΈ° 전에 μš°λ¦¬κ°€ 쀀비물을 챙겨야 되겠죠? κ΅­μ–΄ 여행을 λ– λ‚  μ€€λΉ„λ¬Ό, κ΅μ•ˆμ„ μ–΄λ–»κ²Œ 받을 수 μžˆλŠ”μ§€ μ„ μƒλ‹˜μ΄ μ„€λͺ…을 ν•΄μ€„κ²Œμš”.
(EBS)/(μ΄λΉ„μ—μŠ€) μ΄ˆλ“±μ„ κ²€μƒ‰ν•΄μ„œ λ“€μ–΄κ°€λ©΄μš” 첫화면이 μ΄λ ‡κ²Œ λ‚˜μ™€μš”.
자/ κ·ΈλŸ¬λ©΄μš” μ—¬κΈ° (X)/(μ—‘μŠ€) 눌러주(κ³ μš”)/(κ΅¬μš”). μ €κΈ° (동그라미)/(λ˜₯그라미) (EBS)/(μ΄λΉ„μ—μŠ€) (2μ£Ό)/(이 μ£Ό) λΌμ΄λΈŒνŠΉκ°•μ΄λΌκ³  λ˜μ–΄μžˆμ£ ?
κ±°κΈ°λ₯Ό λ°”λ‘œ κ°€κΈ°λ₯Ό λˆ„λ¦…λ‹ˆλ‹€. 자/ (λˆ„λ₯΄λ©΄μš”)/(눌λ₯΄λ©΄μš”). μ–΄λ–»κ²Œ λ˜λƒ? b/ λ°‘μœΌλ‘œ λ‚΄λ €μš” λ‚΄λ €μš” λ‚΄λ €μš” μ­‰ λ‚΄λ €μš”.
우리 λͺ‡ 학년이죠? μ•„/ (2ν•™λ…„)/(이 ν•™λ…„) 이죠 (2ν•™λ…„)/(이 ν•™λ…„)의 무슨 κ³Όλͺ©? κ΅­μ–΄.
μ΄λ²ˆμ£ΌλŠ” (1μ£Ό)/(일 μ£Ό) μ°¨λ‹ˆκΉŒμš” μ—¬κΈ° κ΅μ•ˆ. λ‹€μŒμ£ΌλŠ” μ—¬κΈ°μ„œ λ‹€μš΄μ„ λ°›μœΌλ©΄ λΌμš”.
이 κ΅μ•ˆμ„ 클릭을 ν•˜λ©΄, μ§œμž”/. μ΄λ ‡κ²Œ κ΅μž¬κ°€ λ‚˜μ˜΅λ‹ˆλ‹€ .이 κ΅μ•ˆμ„ (λ‹€μš΄)/(λ”°μš΄)λ°›μ•„μ„œ 우리 ꡭ어여행을 λ– λ‚  μˆ˜κ°€ μžˆμ–΄μš”.
그럼 우리 μ§„μ§œλ‘œ κ΅­μ–΄ 여행을 ν•œλ²ˆ λ– λ‚˜λ³΄λ„λ‘ ν•΄μš”? κ΅­μ–΄μ—¬ν–‰ 좜발. 자/ (1단원)/(일 단원) 제λͺ©μ΄ λ­”κ°€μš”? ν•œλ²ˆ μ°Ύμ•„λ΄μš”.
μ‹œλ₯Ό μ¦κ²¨μš” μ—μš”. κ·Έλƒ₯ μ‹œλ₯Ό μ½μ–΄μš” κ°€ μ•„λ‹ˆμ—μš”. μ‹œλ₯Ό 즐겨야 λΌμš” 즐겨야 돼. μ–΄λ–»κ²Œ 즐길까? 일단은 λ‚΄λ‚΄ μ‹œλ₯Ό μ¦κΈ°λŠ” 방법에 λŒ€ν•΄μ„œ 곡뢀λ₯Ό ν•  κ±΄λ°μš”.
그럼 μ˜€λŠ˜μ€μš” μ–΄λ–»κ²Œ μ¦κΈΈκΉŒμš”? 였늘 곡뢀할 λ‚΄μš©μ€μš” μ‹œλ₯Ό μ—¬λŸ¬ 가지 λ°©λ²•μœΌλ‘œ 읽기λ₯Ό κ³΅λΆ€ν• κ²λ‹ˆλ‹€.
μ–΄λ–»κ²Œ μ—¬λŸ¬κ°€μ§€ λ°©λ²•μœΌλ‘œ μ½μ„κΉŒ 우리 곡뢀해 보도둝 ν•΄μš”. 였늘의 μ‹œ λ‚˜μ™€λΌ μ§œμž”/! μ‹œκ°€ λ‚˜μ™”μŠ΅λ‹ˆλ‹€ μ‹œμ˜ 제λͺ©μ΄ λ­”κ°€μš”? λ‹€νˆ° λ‚ μ΄μ—μš” λ‹€νˆ° λ‚ .
λˆ„κ΅¬λž‘ λ‹€ν‰œλ‚˜ λ™μƒμ΄λž‘ λ‹€ν‰œλ‚˜ μ–Έλ‹ˆλž‘ μΉœκ΅¬λž‘? λˆ„κ΅¬λž‘ λ‹€ν‰œλŠ”μ§€ μ„ μƒλ‹˜μ΄ μ‹œλ₯Ό 읽어 쀄 ν…Œλ‹ˆκΉŒ ν•œλ²ˆ 생각을 해보도둝 ν•΄μš”."""
inputs = [prefix + sample]
inputs = tokenizer(inputs, max_length=512, truncation=True, return_tensors="pt")
output = model.generate(**inputs, num_beams=3, do_sample=True, min_length=10, max_length=64)
decoded_output = tokenizer.batch_decode(output, skip_special_tokens=True)[0]
result = nltk.sent_tokenize(decoded_output.strip())[0]
print('RESULT >>', result)
RESULT >> κ΅­μ–΄ 여행을 λ– λ‚˜κΈ° 전에 κ΅­μ–΄ 여행을 λ– λ‚  μ€€λΉ„λ¬Όκ³Ό κ΅μ•ˆμ„ μ–΄λ–»κ²Œ 받을 수 μžˆλŠ”μ§€ μ„ μƒλ‹˜μ΄ μ„€λͺ…ν•΄ μ€€λ‹€.
```
# Evalutation Result
- Korean Paper Summarization Dataset(λ…Όλ¬Έμžλ£Œ μš”μ•½)
```
ROUGE-2-R 0.09868624890432466
ROUGE-2-P 0.9666714545849712
ROUGE-2-F 0.17250881441169427
```
- Korean Book Summarization Dataset(λ„μ„œμžλ£Œ μš”μ•½)
```
ROUGE-2-R 0.1575686156943213
ROUGE-2-P 0.9718318136896944
ROUGE-2-F 0.26548116834852586
```
- Korean Summary statement and Report Generation Dataset(μš”μ•½λ¬Έ 및 레포트 생성 데이터)
```
ROUGE-2-R 0.0987891733555808
ROUGE-2-P 0.9276946867981899
ROUGE-2-F 0.17726493110448185
```
# Training
The model was trained with the parameters:
- training arguments
```
Seq2SeqTrainingArguments(
per_device_train_batch_size=8,
per_device_eval_batch_size=8,
auto_find_batch_size=False,
weight_decay=0.01,
learning_rate=4e-05,
lr_scheduler_type=linear,
num_train_epochs=3,
fp16=True)
```
# Model Architecture
```
T5ForConditionalGeneration(
(shared): Embedding(50358, 768)
(encoder): T5Stack(
(embed_tokens): Embedding(50358, 768)
(block): ModuleList(
(0): T5Block(
(layer): ModuleList(
(0): T5LayerSelfAttention(
(SelfAttention): T5Attention(
(q): Linear(in_features=768, out_features=768, bias=False)
(k): Linear(in_features=768, out_features=768, bias=False)
(v): Linear(in_features=768, out_features=768, bias=False)
(o): Linear(in_features=768, out_features=768, bias=False)
(relative_attention_bias): Embedding(32, 12)
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(1): T5LayerFF(
(DenseReluDense): T5DenseGatedActDense(
(wi_0): Linear(in_features=768, out_features=2048, bias=False)
(wi_1): Linear(in_features=768, out_features=2048, bias=False)
(wo): Linear(in_features=2048, out_features=768, bias=False)
(dropout): Dropout(p=0.1, inplace=False)
(act): NewGELUActivation()
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
(1~11): T5Block(
(layer): ModuleList(
(0): T5LayerSelfAttention(
(SelfAttention): T5Attention(
(q): Linear(in_features=768, out_features=768, bias=False)
(k): Linear(in_features=768, out_features=768, bias=False)
(v): Linear(in_features=768, out_features=768, bias=False)
(o): Linear(in_features=768, out_features=768, bias=False)
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(1): T5LayerFF(
(DenseReluDense): T5DenseGatedActDense(
(wi_0): Linear(in_features=768, out_features=2048, bias=False)
(wi_1): Linear(in_features=768, out_features=2048, bias=False)
(wo): Linear(in_features=2048, out_features=768, bias=False)
(dropout): Dropout(p=0.1, inplace=False)
(act): NewGELUActivation()
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
)
(final_layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(decoder): T5Stack(
(embed_tokens): Embedding(50358, 768)
(block): ModuleList(
(0): T5Block(
(layer): ModuleList(
(0): T5LayerSelfAttention(
(SelfAttention): T5Attention(
(q): Linear(in_features=768, out_features=768, bias=False)
(k): Linear(in_features=768, out_features=768, bias=False)
(v): Linear(in_features=768, out_features=768, bias=False)
(o): Linear(in_features=768, out_features=768, bias=False)
(relative_attention_bias): Embedding(32, 12)
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(1): T5LayerCrossAttention(
(EncDecAttention): T5Attention(
(q): Linear(in_features=768, out_features=768, bias=False)
(k): Linear(in_features=768, out_features=768, bias=False)
(v): Linear(in_features=768, out_features=768, bias=False)
(o): Linear(in_features=768, out_features=768, bias=False)
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(2): T5LayerFF(
(DenseReluDense): T5DenseGatedActDense(
(wi_0): Linear(in_features=768, out_features=2048, bias=False)
(wi_1): Linear(in_features=768, out_features=2048, bias=False)
(wo): Linear(in_features=2048, out_features=768, bias=False)
(dropout): Dropout(p=0.1, inplace=False)
(act): NewGELUActivation()
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
(1~11): T5Block(
(layer): ModuleList(
(0): T5LayerSelfAttention(
(SelfAttention): T5Attention(
(q): Linear(in_features=768, out_features=768, bias=False)
(k): Linear(in_features=768, out_features=768, bias=False)
(v): Linear(in_features=768, out_features=768, bias=False)
(o): Linear(in_features=768, out_features=768, bias=False)
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(1): T5LayerCrossAttention(
(EncDecAttention): T5Attention(
(q): Linear(in_features=768, out_features=768, bias=False)
(k): Linear(in_features=768, out_features=768, bias=False)
(v): Linear(in_features=768, out_features=768, bias=False)
(o): Linear(in_features=768, out_features=768, bias=False)
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(2): T5LayerFF(
(DenseReluDense): T5DenseGatedActDense(
(wi_0): Linear(in_features=768, out_features=2048, bias=False)
(wi_1): Linear(in_features=768, out_features=2048, bias=False)
(wo): Linear(in_features=2048, out_features=768, bias=False)
(dropout): Dropout(p=0.1, inplace=False)
(act): NewGELUActivation()
)
(layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
)
)
(final_layer_norm): T5LayerNorm()
(dropout): Dropout(p=0.1, inplace=False)
)
(lm_head): Linear(in_features=768, out_features=50358, bias=False)
)
```
## Citation
- Raffel, Colin, et al. "Exploring the limits of transfer learning with a unified text-to-text transformer." J. Mach. Learn. Res. 21.140 (2020): 1-67.