🤗 KLUE RE(Relation Extraction) Dataset으로 주어진 문장의 지정된 두 Entity의 관계를 추출, 분류하는 Task.
🤗 Public, Private 데이터가 분리된 Leaderboard 평가가 이루어짐.
🤗 하루 10회로 모델 제출 제한

• 'klue/roberta-large' with BiLSTM
• Modify Input format
  • Typed Entity Marker with Punctuation
  • Add Query like Question and Answering
• Augmentation
  • Subject & Object Entity Random Masking
  • AEDA
  • Random Delete
  • Entity swap
• Ensemble
  • Stratified K-Fold & OOF(Out-of-Fold) Prediction
  • K-fold Ensemble via weighted soft voting

• Model docs
  • Developed models
• Augmentation docs
  • Task Adaptive Pre-Training via Back translation
  • EDA
  • AEDA
  • Random masking

pip install -r requirements.txt

# default wandb setting in train.py
run = wandb.init(project= 'klue', entity= 'quarter100', name= f'KFOLD_{fold}_{args.wandb_path}')

python train.py

Models are saved in "./best_model/".

python inference_fold.py

Prediction csv files are saved in "./prediction".

python vote.py

Ensemble result is saved in "./prediction/submission_fold_total.csv".