Global and Local Critical Policy Learning for Abstractive Summarization

N.B. Includes a Pytorch implementation of Get To The Point: Summarization with Pointer-Generator Networks Base version reference: pointer_summerizer

Training

• To train an initial Seq-2-Seq Network run the training_ptr_gen/train_pytorch.py file
• To train the Network using the Policy Gradient run the seqGAN/main.py file (be sure to point to your pretrained Seq2Seq model if you want good results). The default reward level is the global sentence level reward.
• Pass --word-level to train with local word level rewards
• Pass --combined to train with the combined rewards
• Adjust the data_util/config.py as necessary to adjust paths and training settings
• N.B. You may need to run export PYTHONPATH=$PYTHONPATH:'pwd' in order to properly configure the references

Data:

• Follow data generation instruction from https://github.com/abisee/cnn-dailymail

Dependencies

• Requires python 3.x and pytorch 0.4
• To install the dependencies run pip install -r requirements.txt
• In order to evaluate results using the ROUGE metric please follow the instructions below (a working PERL installation is required)

How to Install PyRouge

---

1. Clone old PyRouge Repository into separate directory git clone https://github.com/andersjo/pyrouge.git pyrouge_ancient
2. Clone new PyRouge Repository git clone https://github.com/bheinzerling/pyrouge.git
3. Run pip install pyrouge
4. Set the Rouge-1.5.5 Installation Directory (N.B. This needs to be an absolute path!!!) pyrouge_set_rouge_path <absolute_path_of_current_directory>/pyrouge_ancient/tools/ROUGE-1.5.5/
5. Run cd pyrouge
6. Check to see if it worked by running python -m pyrouge.test

How to install python pyrouge

---

1. Follow the install instructions of git clone https://github.com/pltrdy/rouge

Evaluation of saved models

• Note: For training passing the initial checkpoint is optional argument, but for evaluation and decoding passing the checkpoint is must.
• For evaluation and decoding:

python training_ptr_gen/eval.py <absolute/path/to/model/checkpoint>
python training_ptr_gen/decode.py <absolute/path/to/model/checkpoint>