QAnet Chinese Reading Comprehension

demo 畫面 https://github.com/sinlin0908/QAmodel_Demo

• QAnet Chinese Reading Comprehension

• prepro file

• embedding dat set

• Config

• Reference

• Environment

• Requirement

• Data Set

• Modify

• Usage

• Performance

• Tensorbord

• Comparison Chart

Reference

• QAnet by google

• QAnet Code by NLP_learn

• Data Set by Delta

• Embedding data set by Chinese-Word-Vectors

Environment

• OS: Ubuntu 18.04 LTS

• GPU: GTX 1080Ti 11G

• CPU: i7-4770

• RAM: 16G

Requirement

• Python 3.6

• NumPy

• tqdm

• Tensorflow>=1.5

• Jieba

• opencc

• bottle

Data Set

• train set : 26936 questions

• dev set : 3524 questions

• test 3493 : questions

Modify

prepro file

• word token: use jieba.cut(context,cut_all=False)

• _getword() delete word.lower(), word.capitalize(), word.upper()

embedding dat set

• use 1292607 words 300d embedding data set

• use 14082 characters 300d character embedding data set

Config

• word size: 1292607

• hidden size: 128

• num_head: 8

• batch size: 12

• char_emb_size : 300d

• pretrain_char -> True

Usage

python config.py --mode prepro

train

python config.py --mode train

test

python config.py --mode test

demo:

python config.py --mode demo

Performance

• F1: score 70.0496230556

• EM: 70.0257658173

• cost: 6 hours

• use GPU memory : 9.4G

Tensorbord

tensorboard --logdir=./

Comparison Chart

notice: character embedding has a little effect