Extending the Bi-directional Attention Flow Model with "No Answer"


  • Python (developed on 3.5.2. Issues have been reported with Python 2!)
  • tensorflow (deep learning library, verified on r0.11)
  • nltk (NLP tools, verified on 3.2.1)
  • tqdm (progress bar, verified on 4.7.4)


  • <run name> calls an internal script ( that changes our tab-delimited format to SQuAD's JSON format. It then performs any necessary preprocessing for the BiDAF model.
  • <run name> runs the training procedure.
  • <run name> runs the testing procedure, and yields an answer file in out/basic/<run name>/test-#####.json
  • python <test set> <answer file> reads the test set and the model's answers, and returns the F1 score broken down by different factors.