Dynamic Coattention Networks for SQuAD Question Answering System

This is a tensorflow (Python2 with Tensorflow 0.12.1) implementation of a variant of the dynamic coattention networks for the question answering system based on the SQuAD dataset. Here is the report.

Download Dataset

sh get_started.sh
python2 qa_data.py --glove_dim [dim]

[dim] denotes the GloVe word embedding size, and can be 50, 100, 200, or 300. They have been pretrained on Wikipedia 2014 and Gigaword 5, with a vocabulary size of 400k.

Training

python2 train.py

Generate Answers

qa_answer.py is used to generate answers from a trained model. It will take a JSON file as input and output a JSON file containing the answers. The format of JSON file can be seen in the downloaded dataset.