Leveraging Information Bottleneck for Scientific Document Summarization

EMNLP 2021 Findings paper: Leveraging Information Bottleneck for Scientific Document Summarization

Requirements

IBsumm is implemented using Python3.7 with dependencies specified in requirements.txt. To download all the required packages, use the code block below:

!pip install -r requirements.txt
!pip install transformers==3.4.0
!pip install pytorch-transformers
!pip install tensorflow==2.3.0
!pip install urllib3==1.25.10
!pip install fastai==1.0.61
!pip install --upgrade numpy scipy pandas
import nltk
nltk.download(['punkt','stopwords','wordnet','omw-1.4'])

For global labels computed using pre-trained Longformer, please download here. Place the downloaded pickle file to the "models" directory.

Extractive Summarization

Place the source document file in the source directory. The source document file should contain the source text in each line.

Toy Example

A simple model run based on the first 5 articles of COVID-19 dataset.

python main.py --input_path 'source' --input_file 'toy_test.txt' --output_path 'output' --output_file 'toy_outputs.txt' --model_path 'Longformer_global_label.pkl'

arXiv

python main.py --input_path 'source' --input_file 'arxiv_test.txt' --output_path 'output' --output_file 'arxiv_outputs.txt' --model_path 'Longformer_global_label.pkl'

COVID-19

python main.py --input_path 'source' --input_file 'covid19_test.txt' --output_path 'output' --output_file 'covid19_outputs.txt' --model_path 'Longformer_global_label.pkl'

PubMed

python main.py --input_path 'source' --input_file 'pubmed_test.txt' --output_path 'output' --output_file 'pubmed_outputs.txt' --model_path 'Longformer_global_label.pkl'

Citation

Please consider citing our work:

@inproceedings{ju2021leveraging,
  title={Leveraging Information Bottleneck for Scientific Document Summarization},
  author={Ju, Jiaxin and Liu, Ming and Koh, Huan Yee and Jin, Yuan and Du, Lan and Pan, Shirui},
  booktitle={Findings of the Association for Computational Linguistics: EMNLP 2021},
  pages={4091--4098},
  year={2021}
}