t5-base-korean-summarization
This is T5 model for korean text summarization.
Finetuned based on 'paust/pko-t5-base' model.
Finetuned with 3 datasets. Specifically, it is described below.
Usage (HuggingFace Transformers)
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 = """
μλ
νμΈμ? μ°λ¦¬ (2νλ
)/(μ΄ νλ
) μΉκ΅¬λ€ μ°λ¦¬ μΉκ΅¬λ€ νκ΅μ κ°μ μ§μ§ (2νλ
)/(μ΄ νλ
) μ΄ λκ³ μΆμλλ° νκ΅μ λͺ» κ°κ³ μμ΄μ λ΅λ΅νμ£ ?
κ·Έλλ μ°λ¦¬ μΉκ΅¬λ€μ μμ κ³Ό 건κ°μ΄ μ΅μ°μ μ΄λκΉμ μ€λλΆν° μ μλμ΄λ λ§€μΌ λ§€μΌ κ΅μ΄ μ¬νμ λ λ보λλ‘ ν΄μ.
μ΄/ μκ°μ΄ λ²μ¨ μ΄λ κ² λλμ? λ¦μμ΄μ. λ¦μμ΄μ. 빨리 κ΅μ΄ μ¬νμ λ λμΌ λΌμ.
κ·Έλ°λ° μ΄/ κ΅μ΄μ¬νμ λ λκΈ° μ μ μ°λ¦¬κ° μ€λΉλ¬Όμ μ±κ²¨μΌ λκ² μ£ ? κ΅μ΄ μ¬νμ λ λ μ€λΉλ¬Ό, κ΅μμ μ΄λ»κ² λ°μ μ μλμ§ μ μλμ΄ μ€λͺ
μ ν΄μ€κ²μ.
(EBS)/(μ΄λΉμμ€) μ΄λ±μ κ²μν΄μ λ€μ΄κ°λ©΄μ 첫νλ©΄μ΄ μ΄λ κ² λμμ.
μ/ κ·Έλ¬λ©΄μ μ¬κΈ° (X)/(μμ€) λλ¬μ£Ό(κ³ μ)/(ꡬμ). μ κΈ° (λκ·ΈλΌλ―Έ)/(λ₯κ·ΈλΌλ―Έ) (EBS)/(μ΄λΉμμ€) (2μ£Ό)/(μ΄ μ£Ό) λΌμ΄λΈνΉκ°μ΄λΌκ³ λμ΄μμ£ ?
κ±°κΈ°λ₯Ό λ°λ‘ κ°κΈ°λ₯Ό λλ¦
λλ€. μ/ (λλ₯΄λ©΄μ)/(λλ₯΄λ©΄μ). μ΄λ»κ² λλ? b/ λ°μΌλ‘ λ΄λ €μ λ΄λ €μ λ΄λ €μ μ λ΄λ €μ.
μ°λ¦¬ λͺ νλ
μ΄μ£ ? μ/ (2νλ
)/(μ΄ νλ
) μ΄μ£ (2νλ
)/(μ΄ νλ
)μ λ¬΄μ¨ κ³Όλͺ©? κ΅μ΄.
μ΄λ²μ£Όλ (1μ£Ό)/(μΌ μ£Ό) μ°¨λκΉμ μ¬κΈ° κ΅μ. λ€μμ£Όλ μ¬κΈ°μ λ€μ΄μ λ°μΌλ©΄ λΌμ.
μ΄ κ΅μμ ν΄λ¦μ νλ©΄, μ§μ/. μ΄λ κ² κ΅μ¬κ° λμ΅λλ€ .μ΄ κ΅μμ (λ€μ΄)/(λ°μ΄)λ°μμ μ°λ¦¬ κ΅μ΄μ¬νμ λ λ μκ° μμ΄μ.
κ·ΈλΌ μ°λ¦¬ μ§μ§λ‘ κ΅μ΄ μ¬νμ νλ² λ λ보λλ‘ ν΄μ? κ΅μ΄μ¬ν μΆλ°. μ/ (1λ¨μ)/(μΌ λ¨μ) μ λͺ©μ΄ λκ°μ? νλ² μ°Ύμλ΄μ.
μλ₯Ό μ¦κ²¨μ μμ. κ·Έλ₯ μλ₯Ό μ½μ΄μ κ° μλμμ. μλ₯Ό μ¦κ²¨μΌ λΌμ μ¦κ²¨μΌ λΌ. μ΄λ»κ² μ¦κΈΈκΉ? μΌλ¨μ λ΄λ΄ μλ₯Ό μ¦κΈ°λ λ°©λ²μ λν΄μ 곡λΆλ₯Ό ν 건λ°μ.
κ·ΈλΌ μ€λμμ μ΄λ»κ² μ¦κΈΈκΉμ? μ€λ 곡λΆν λ΄μ©μμ μλ₯Ό μ¬λ¬ κ°μ§ λ°©λ²μΌλ‘ μ½κΈ°λ₯Ό 곡λΆν κ²λλ€.
μ΄λ»κ² μ¬λ¬κ°μ§ λ°©λ²μΌλ‘ μ½μκΉ μ°λ¦¬ 곡λΆν΄ 보λλ‘ ν΄μ. μ€λμ μ λμλΌ μ§μ/! μκ° λμμ΅λλ€ μμ μ λͺ©μ΄ λκ°μ? λ€ν° λ μ΄μμ λ€ν° λ .
λꡬλ λ€νλ λμμ΄λ λ€νλ μΈλλ μΉκ΅¬λ? λꡬλ λ€νλμ§ μ μλμ΄ μλ₯Ό μ½μ΄ μ€ ν
λκΉ νλ² μκ°μ ν΄λ³΄λλ‘ ν΄μ."""
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΄ μ¬νμ λ λκΈ° μ μ κ΅μ΄ μ¬νμ λ λ μ€λΉλ¬Όκ³Ό κ΅μμ μ΄λ»κ² λ°μ μ μλμ§ μ μλμ΄ μ€λͺ
ν΄ μ€λ€.
Evalutation Result
- Korean Paper Summarization Dataset(λ
Όλ¬Έμλ£ μμ½)
ROUGE-2-R 0.09868624890432466 ROUGE-2-P 0.9666714545849712 ROUGE-2-F 0.17250881441169427
- Korean Book Summarization Dataset(λμμλ£ μμ½)
ROUGE-2-R 0.1575686156943213 ROUGE-2-P 0.9718318136896944 ROUGE-2-F 0.26548116834852586
- Korean Summary statement and Report Generation Dataset(μμ½λ¬Έ λ° λ ν¬νΈ μμ± λ°μ΄ν°)
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.
- Downloads last month
- 625,962