=========================================

This package contains code for the Exploration of various deep neural networks for Question Answering.

With different configs for different models in config directory, you will be able to train a model on the SQuAD dataset.

To train a model you should run python train.py name_of_config (no .py extension for config name). Before that you should get the SQuAD data and tokenize it using the tokenize_data function in utils.py file. This file also contains functions to create a vocabulary, compute length coverege, etc.

All the models are in the model directory, from the basic seq2seq model with attention to the Match-LSTM model with pointer networks.

Fuel is being used for the data pipeline. Iterator and Dataset classes for SQuAD and CNN/DailyMotion datasets are available in data.py as well as a Toy Dataset for debugging purposes.

Evaluation extensions for models in lmu_extensions.py are implemented to be used with Blocks' extensions. For seq2seq models average recall, average precision, macro F1, average F1 and exact match accuracies are reported.

The heatmap shows the attention paid to each token of the question at each step of encoding the paragraph.

As you can see the SQuAD dataset contains some answers with high lengths which will evidently make sequence to sequence solutions challenging for this task.

(hotizontal axis is the length and vertical axis is the acc coverage)

We would like to thank the developers of Theano, Blocks and Fuel at MILA for their excellent work.