/CNN-in-Answer-selection

WikiQA,复现论文《APPLYING DEEP LEARNING TO ANSWER SELECTION: A STUDY AND AN OPEN TASK》

Primary LanguagePythonMIT LicenseMIT

WikiQA on QACNN

复现论文《APPLYING DEEP LEARNING TO ANSWER SELECTION: A STUDY AND AN OPEN TASK》

本项目采取了论文中最好的模型进行实验,数据集采用WikiQA,后期会上传insuranceQA的实验结果
模型图如下:
model
实验结果

Model CNN share Dropout Parameters Margin Epoch MAP MRR
QACNN No 0.5 2115200 0.5 100 0.655 0.673
QACNN Yes 0.5 481664 0.5 100 0.684 0.697
QACNN Yes 0.5 481664 0.25 100 0.668 0.674
QACNN Yes 0.5 481664 0.2 100 0.690 0.695

LossPairwise Loss

有时间就会更新QA实验,有兴趣的同学可以follow一下,也欢迎Fork和Star!
留言请在Issues或者email xiezhengwen2013@163.com