• Python >= 3.6
• Tensorflow >= 1.6 (self-compiled TF-gpu is recommended!)
• gputil >= 1.3.0
• GPU

For dual learning run:

python grid_search.py daanet 

For QA-only model (corresponds to mono in the experiment) run:

python grid_search.py monoqa 

For QG-only model (corresponds to mono in the experiment) , run :

python grid_search.py monoqg 

All hyperparameters used in the paper are stored in default.yaml:

You can change the data path and save dir in grid.yaml

Evaluation on the dev set is automatically done after each epoch.

To do evaluation manually,

python app.py evaluate ~/save/models/DDMM-HHMMSS/default.yaml

, where ~/save/models/DDMM-HHMMSS/default.yaml is the saved yaml config of model DDMM-HHMMSS. It is created during the training procedure. It automatically loads the parameters from the best epoch (or fallback to the last epoch) to the model.

python app.py train ~/save/models/DDMM-HHMMSS/default.yaml

It will load the best (or last) model so far and conducts incremental training.

Selected outputs from DAANET and mono. Yellow text is question-related; green text is answer-related.

Question-Context and Answer Context attention matrices