zzoy / machinecomprehension-1

Machine comprehension for recent papers using pytorch

MachineComprehension

Introduction

Pytorch implementation of machine comprehension papers for SQuAD v1.1

BIDAF (Minjoon Seo et al.,2016) https://arxiv.org/abs/1611.01603
QANet (Adams Wei Yu et al.,2018) https://arxiv.org/abs/1804.09541
Ensemble model for BIDAF and QANet

Structure

• preproc.py: dataset preprocessing and build training features
  
• config.py: determine which model to train and hype-parameters setting
  
• evaluate.py: evaluate script
  
• main.py: program entry
  
• models/qanet.py: QANet model
  
• models/bidaf.py: BIDAF model
  
• models/ensemble.py QANet and BIDAF ensemble
  

Result

Result on dev dataset

	QANet	BIDAF	Ensemble
F1	76.3	74.1	77.6
EM	67.5	63.3	68

Difference with papers

• Context length is set to 300 due to limit of memory
• Char embedding doesn't be connected with convolution layer
• Difference hype parameters setting

Todo

• Achieve the result in paper
• Reduce memory cost
• Complete R-net model
• Do the ensemble