Description

It is the implementation of seq2seq with attention model from [Neural Question Generation from Text: A Preliminary Study] paper.

Results almost the same as in the paper.

Environment

• Python 3.6.0
• tensorflow 1.1.0
• Reverse-SQuAD dataset, download by the link. This corpus should be placed in the root of project (./redistribute/).
• Install google/seq2seq - goto seq2seq dir and do steps from README.md.

All packages are listed in requirements.txt.

Reproduce results

0. Setup the environment.
1. Run: bin/setup
2. Run: bin/seq2seq_with_att
3. Goto tensorboard and see BLEU in dev set.
4. To check in test set modify and run bin/seq2seq_with_att_eval.

Results

I've got 3.03 BLEU on dev set.

There is open issue with model inference. More information here: google/seq2seq#214.