/BiDAF

Re-implementation of Bidirectional Attention Flow for Machine Comprehension (ICLR 2017)

Primary LanguagePython

BiDAF

Re-implementation of Bidirectional Attention Flow for Machine Comprehension (ICLR 2017)

Requirements

  • python=3.7.0
  • pytorch=1.1.0
  • torchtext
  • tensorboardX
  • tqdm
  • nltk

Train model

  • Squad1.1 where every question has its answer.
MODEL_NAME="model"
python main.py --model_name ${MODEL_NAME}\
    --GPU 0\
    --train_batch_size 60\
    --train_file train-v1.1.json\
    --dev_batch_size 100\
    --dev_file dev-v1.1.json

Train model with unanswerable questions (TBD)

  • Squad 2.0 where some questions don't have their answers within the passages.
MODEL_NAME="model"
python main.py --model_name ${MODEL_NAME}\
    --GPU 0\
    --train_batch_size 60\
    --train_file train-v2.0.json\
    --dev_batch_size 100\
    --dev_file dev-v2.0.json\
    --no_answer True

Evaluatation

checkout dataset EM F1
edca354e8b74d895c6f5a363f31aefab267dbf89 SQuAD1.1 65.13 75.53
edca354e8b74d895c6f5a363f31aefab267dbf89 SQuAD2.0 (no_answer False) 32.87 38.31
SQuAD2.0 ? ?