English-to-Korean Transliterator | 영한 음역기

▶ README: ENGLISH | 한국어

This project focuses on the English-to-Korean transliteration task (e.g. english -> 잉글리시). Specifically, this project accelerates inference speed and reduces the large model size (~1.2GB) of the previous MT5-based Transliterator.

• Highlights

  1. Providing lightweight model size (~400MB) with faster, accurate transliteration results.
     • Please check performance comparisons below.
  2. Applying LoRA to the MarianMT translation model.
     • The corresponding fine-tuned model is available in HuggingFace.

• Performance Comparisons

  

How to Start

1. Install dependencies (Assuming PyTorch is already installed):

   pip install -r requirements.txt

2. Clone Repository:

   git clone https://github.com/feVeRin/enko_transliterator.git

How to Use

1. Transliteration (w/ pre-trained model)

   from transliteration import Transliterator
   
   model = Transliterator.from_pretrained('feVeRin/enko-transliterator')
   result = model.transliterate('LORA IS ALL YOU NEED')
   print(result)  # 로라 이즈 올 유 니드

2. Model Training (from scratch)

   • Training is linked to wandb. If not necessary, add report_to=None to the train() function.

   from train import LoRATrainer
   
   trainer = LoRATrainer()
   trainer.set_lora(r=16, alpha=32, dropout=0.1)
   train_dataset, val_dataset = trainer.data_split('./data/data.txt', 0.2)
   trainer.train(train_dataset, val_dataset)

References

• This project utilized a dataset of the EngtoKor-Transliterator.
• Opus-hplt-EN-KO was used as the base model for applying LoRA.