search

matthew-z / R-net

R-net in PyTorch, with ELMo
MIT License
199 stars 43 forks source link

question about accuracy #4

Closed chenjiaxiang closed 6 years ago

chenjiaxiang commented 6 years ago

hello，i don't find a email address on your homepage，so leave an issue. I ran your r-net code, and i found the best accuracy is or so 0.66, the result is not as good as result they reported in the paper, you had pointed out that you directly used glove in char embedding. But the gap shouldn't be so big, i want to know the best result on your machine. Thanks a lot. You can reply here or send email to me. cjx123@mail.ustc.edu.cn

matthew-z commented 6 years ago

Sorry, I have not found a way to reproduce the paper’s result. 0.66 is also the result on my machine.

chenjiaxiang commented 6 years ago

Thanks, I will try to add char level embedding and run the code

发自我的 iPhone

在 2017年12月18日，21:52，EviscerateZ notifications@github.com 写道：

Sorry, I have not found a way to reproduce the paper’s result. 0.66 is also the result on my machine.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

matthew-z commented 6 years ago

Actually, I have added convolutional char embedding in my local dev branch. The performance improved only a little (1%). In their ACL paper, they also reported that the performance decreased by 0.9% F1 score without char embedding.

chenjiaxiang commented 6 years ago

I know.thanks lot

发自我的 iPhone

在 2017年12月20日，10:44，EviscerateZ notifications@github.com 写道：

Actually, I have added convolutional char embedding in my local dev branch. The performance improved only a little (1~2%).

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.