Skip to content

Concept-Pointer-Network-for-Abstractive-Summarization

Notifications You must be signed in to change notification settings

wprojectsn/codes

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

31 Commits
 
 
 
 

Repository files navigation

codes

Pytorch implementation for paper of "Concept Pointer Network for Abstractive Summarization".

pytorch 0.4 with python 2.7

How to train?
before train
  • Our conceptual vocabulary in code/data/vocabulary/concept_vocab,the number of conceptual words needs to be set before train.
  • You need to change some path and parameters in code/codespace/data_util/config.py before train.
Cross-Entropy Object Function
  • python train.py None or python train.py (position of model parameters)
  • When you use Cross-Entropy Object Function train the model,set RL_train = False and DS_train = False in code/codespace/data_util/config.py
reinforcement learning(RL)
  • Before use RL train the model, you need use Cross-Entropy Object Function train the model and set the Cross-Entropy train times in code/codespace/data_util/config.py, finally the model will automatically use reinforcement learning to train the model when the train times exceed the Cross-Entropy train times.
  • When you use RL train the model,set RL_train = True and DS_train = False in code/codespace/data_util/config.py
Distant Supervision(DS)
  • Before use DS train the model, you need use Cross-Entropy Object Function train the model and Retain model parameters,then use the command “python train.py (position of model parameters)” train the model.
  • When you use DS train the model,set DS_train = True and RL_train = False in code/codespace/data_util/config.py
How to test?
  • python decode.py (position of model parameters)

other

About

Concept-Pointer-Network-for-Abstractive-Summarization

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages