/BERT-Question-Generation

A Recurrent BERT-based Model for Question Generation (BERT Text Generation)

Primary LanguagePython

BERT-Question-Generation

Try to implement a paper A Recurrent BERT-based Model for Question Generation (EMNLP 2019)

Score (SQuAD 73K, paragraph level context)

In Original Paper

Model BLEU1 BLEU2 BLEU3 BLEU4 METEOR ROUGE-L
BERT-HLSQG 49.73 34.60 26.13 20.33 23.88 48.23

My Model (Epoch 5, bert-base-uncased)

Model BLEU1 BLEU2 BLEU3 BLEU4 METEOR ROUGE-L
BERT-HLSQG 52.26 33.19 22.48 15.64 22.66 45.48

Quick start

install package

pip install -r requirements.txt

Train

bash start.sh ## if you want to save log in nohup.out
# or
python train.py

Inference

python inference.py

Scoring

## setup scorer
python setup_scorer.py

## make srt-text, tgt-test file

## evaluation
python nqg/qgevalcap/eval.py \
  --src ./data/squad_nqg/src-test.txt \
  --tgt ./data/squad_nqg/tgt-test.txt \
  --out ./data/squad_nqg/predict_squad.txt

Reference

https://github.com/voidful/BertGenerate