yfqiu98 / graphaugmentedsum

Knowledge Graph-Augmented Abstractive Summarization with Semantic-Driven Cloze Reward

Knowledge Graph-Augmented Abstractive Summarization with Semantic-Driven Cloze Reward

Code for ACL2020 paper: Knowledge Graph-Augmented Abstractive Summarization with Semantic-Driven Cloze Reward

We are not allowed to share data/outputs on New York Times Dataset. If you need data/outputs on New York Times Dataset, please email me with your license and we're glad to share our processed data/outputs on NYT dataset for research purpose.

My permenant email address: [email protected]

How to train our model

I. our processed data with constructed graphs can be found here:

https://drive.google.com/open?id=1ccja3oyWIJIm91EiG-NJPFNb4Eg1pOmO

our processed cloze questions can be found here:

https://drive.google.com/open?id=16aPmfT9Gurjhg1uLeVAUTL7fTc6TO42W

our best model can be found here:

https://drive.google.com/open?id=19HeT3rr2mzvEx82arrvpSVOBM_JNeRzo

our trained cloze model can be found here:

https://drive.google.com/open?id=1SxIitGBuPmOfKPHQ21LIX_OJ1RUpHpsk

our best system results/reference can be found here:

https://drive.google.com/open?id=1SRLCVb-YtCzL5NgI76CXby_Oc_MczYjk
https://drive.google.com/open?id=1uXn-dyN4KH4LYzKsCDCVnvRDbbR-lAAV

II. To train our best model:

0. specify data path
   export DATA=[path/to/decompressed/data]

1. train our model with ML objective

python train_abstractor.py --batch 32 --max_input 512 --bert --docgraph(--paragraph for SegGraph extension) --path [path/to/ml/model]

2. train our model with our cloze reward

python train_abstractor_rl.py --abs_dir [path/to/ml/model] --docgraph(--paragraph for SegGraph extension) --batch 32 --max_art 512 --reward_model_dir [/path/to/cloze/model] --reward_data_dir [/path/to/cloze/data/questions] --path [/path/to/best/model]

3. decode

python decode_abs.py --abs_dir [/path/to/best/model] --test --reverse --docgraph(--paragraph)  --gpu_id 0 --path [/path/to/results]

4. evaluate ROUGE

export ROUGE=[/path/to/ROUGE 1.5.5]

python evaluate_full_model.py --decode_dir [/path/to/results] --rouge

5. evaluate QA

python eval_cloze_model.py --system_path [/path/to/results] --data_path [/path/to/cloze/data/questions]  --model_dir [/path/to/cloze/model]

III. To train our multiple choice QA model

python train_roberta_multiple_choice.py --path [/path/to/cloze/data/training] --save_path [/path/to/cloze/model]

Dependencies

• Python 3 (tested on python 3.6)
• PyTorch
  • with GPU and CUDA enabled installation (though the code is runnable on CPU, it would be way too slow)
• gensim
• cytoolz
• tensorboardX
• pyrouge (for evaluation)
• transformers