CitationSum

This code is for the implementation of CitationSum The implementation is for

Some codes are from PreSum:https://github.com/nlpyang/PreSumm and CGSum: https://github.com/ChenxinAn-fdu/CGSum

Python version: This code is in Python3.6

Package Requirements: torch==1.1.0 pytorch_transformers tensorboardX multiprocess pyrouge

Data Preparation For SSN

Step 1 Download

Download and unzip the SSN (including inductive and transductive) from here.

Step 2. Format to PyTorch Files

python preprocess.py -mode format_cite -raw_path RAW_PATH -save_path BERT_DATA_PATH  -mode inductive -lower -n_cpus 8 -log_file ../logs/preprocess.log

• RAW_PATH is the directory containing raw files (../inductive), BERT_DATA_PATH is the target directory to save the generated binary files (../bert_data)

PubMedCite Dataset

The citation graph records of our proposed PubMedCite dataset are in PubMedCite.zip. Concerning the different licences of the documents in the dataset, we are unable to release the full articles with the citation graph. Please consider using the official API of PubMed to fetch full data.

Model Training

Abstractive Setting

python train.py  -task abs -mode train -bert_data_path BERT_DATA_PATH -dec_dropout 0.2  -model_path MODEL_PATH -sep_optim true -lr_bert 0.002 -lr_dec 0.2 -save_checkpoint_steps 2000 -batch_size 120 -train_steps 200000 -report_every 50 -accum_count 5 -use_bert_emb true -use_interval true -warmup_steps_bert 20000 -warmup_steps_dec 10000 -max_pos 512 -visible_gpus 0,1,2,3  -log_file ../ssn.log

Model Evaluation

SSN

 python train.py -task abs -mode validate -batch_size 3000 -test_batch_size 500 -bert_data_path BERT_DATA_PATH -log_file ../logs/val_ssn -model_path MODEL_PATH -sep_optim true -use_interval true -visible_gpus 1 -max_pos 640 -max_length 300 -alpha 0.95 -min_length 130 -result_path ../logs/