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

Our concept word come from https://concept.research.microsoft.com/Home/API
Part of our code references https://github.com/atulkum/pointer_summarizer