T5 Machine Translation: English ↔️ Korean

Result

	BLEU Score	Translation Result
Korean ➡️ English	45.148	KE-T5-Ko2En-Base Inference Result
English ➡️ Korean	-

• Evaluation script is on metric.py
• Korean ➡️ English Result evaluated on 553500 sentence pairs which are disjoint from the train set.

How to Use

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Korean -> English Machine Translation
tokenizer = AutoTokenizer.from_pretrained("QuoQA-NLP/KE-T5-Ko2En-Base")
model = AutoModelForSeq2SeqLM.from_pretrained("QuoQA-NLP/KE-T5-Ko2En-Base")

# English -> Korean Machine Translation
tokenizer = AutoTokenizer.from_pretrained("QuoQA-NLP/KE-T5-En2Ko-Base")
model = AutoModelForSeq2SeqLM.from_pretrained("QuoQA-NLP/KE-T5-En2Ko-Base")

• For batch translation, please refer to inference.py.
  • P100 16GB supports inferencing of 250 pairs per batch on device.
  • A100 40GB supports inferencing of 600 pairs per batch on device.
• For single sentence translation, please refer to inference_single.py.

References

• 🔗 Dataset specification
• Translation Example
• Summarization Example
• Deep Encoder Shallow Decoder