-
This is an Implementation of HRCA and HRCA+,which are proposed in the paper:HRCA+: Advanced Multiple-choice Machine Reading Comprehension Method.I also add pdf in this repo.This project is for graduatation design of Tongji University.
-
My Implementation is based on another repo:
https://github.com/pfZhu/duma_code,for the reason that both DUMA and HRCA use mulit-head Attention.
-
git clone my repo
-
run in Linux:
pip install -r requirements.txt -
run in Linux:
sh run_dream.sh
- The number of layers of HRCA and HRCA+ are 4.
- Currently the code is using HRCA+,if you want to use HRCA,just modify line 70 in 'run_multiple_choice.py'
- Some python packages in 'requirements.txt' are not used,however I did not check them out,so if you meet failure in pip installing,perhaps you could try to DELETE it in 'requirements.txt'.
- I do not use fp16 because I failed to install nvidia apex.If you want to use fp16,please refer to the REAMDE.md of
https://github.com/pfZhu/duma_code.
-
It's pity that my results are not as good as the results of the paper,which I think is because the implements of HRCA and HRCA+ are not perfect and the parameters could be adjusted better.
-
These are my results on a single V100 GPU:
Model acc of DREAM learning rate epoches batch_size Albert_base+HRCA 66.7 2e-5 3 2 Albert_base+HRCA+ 67.8 8e-6 3 2
- If you would like to contact me,please send email to 1951444@tongji.edu.cn. Both Chinese and English are OK.
- Please STAR this repo if you think it is helpful.Thanks.