Introduction

The RCZoo project is a toolkit for reading comprehension model. It contains the PyTorch reimplement of multiple reading comprehension model.
All the models are trained and tested on the SQuAD v1.1 dataset, and reach the performance in origin papers.

Dependencies

python 3.5
Pytorch 0.4
tqdm

performance

We train each model on train set for 40 epoch, and report the best performance on dev set.

Model	Exact Match	F1
Rnet	69.25	78.97
BiDAF	70.47	79.90
documentqa	71.47	80.84
DrQA	68.39	77.90
QAnet	...	...
SLQA	67.09	76.67
FusionNet	68.27	77.79

Current progress

Rnet

• training
• performance
• predicting scripts
  some different in the Dropout Layer

BiDAF

• training
• performance
• predicting scripts
  The bi-attention in BiDAF does not work fin, and I introduce the co-attention in DCN paper. The final results is better than that in origin paper

documentqa

• training
• performance
• predicting scripts

DrQA

borrow from origin code

• training
• performance
• predicting scripts

QAnet

• training
• performance
• predicting scripts

SLQA

• training
• performance
• predicting scripts
  no elmo contextualized embedding

FusionNet

• training
• performance
• predicting scripts
  no CoVe embedding

Usage

• run sh download.sh to download the dataset and the glove embeddings.
• run sh train_xxx.sh to start the train process. Dring the train process, model will be evaluated on dev set each epoch.

acknowledgement

some code are borrowed from DrQA, a cool project about reading comprehension.

TODO:

• Recognizing unanswerable question for SQuAD, add new type of loss function to accommodate unanswerable question
• Processing multiple passage reading comprehension. Related datasets include TriviaQA, SearchQA, QuasarT