Question Answering using character-level RNN over babi (FAIR dataset) and SQUAD (Stanford dataset)

The NLP community has made great progress on open-domain question answering, but the systems still struggle to answer complex questions over a large collection of text. This project is a small effort to present an efficient and explainable method for Question Answering using character-level RNNs over the SQUAD and BABI Dataset.

Name	Description
/babi	babi dataset and utilities
/squad	squad dataset and utilities
/qrn	contains the QRN cell
char2word.py	Char2Word-only module (on bAbI dataset)
qrn.py	implementation of the QRN model (on bAbI dataset)
char2word_qrn.py	implementation of the QRN model w/ Char2Word module (on bAbI dataset)
squad.py	implementation of the QRN model w/ Char2Word module (on SQuAD dataset)

Files :

• babi_formatting.py - processes the dataset for character-level handling.
• /datasets - repository for datasets
• /model/dataset.py - loads the data
• /model/model.py - constructs model_fn that will be fed into the Estimator in main.py
• main.py - runs TensorFlow instances to train and evaluate the model
• run.sh - runs main.py given a dataset and a seed
• char2word.py - contains the Char2Word block

Usage (with bAbI dataset):

Query-Reduction Network without Char2Word:

python qrn.py

Query-Reduction Network with Char2Word:

python char2word_qrn.py

Usage (with SQuAD dataset):

Query-Reduction Network with Char2Word:

python squad.py