SCAR: Sentence Compression using Autoencoders for Reconstruction

This is our Keras implementation of the summarization methods described in Sentence Compression using Autoencoders for Reconstruction. It features linkage loss which helps drop inferable words, in turn bringing out content aware summary for a sentence.

Table of Contents

• Requirements
• Quickstart
• Cite

Requirements

virtualenv venv
source venv/bin/activate
pip install -r requirements.txt

Training datasets:

• Gigaword
• Glove embeddings

Evaluate scores on the DUC2003/DUC2004 datasets.

Quickstart

Step 1: Preprocess the dataset

Place the following files in the data directory:

• glove.42B.300d.txt
• train.article.txt
• valid.article.filter.txt

python preprocess.py $expNo$ 
(Example: python preprocess.py 6)

Update no_of_steps and no_of_steps_valid in config.json based on the output (Training/Validation steps/epoch) of the above script.

Step 2: Train the model

Create exp$expNo$ folder (Example: exp6) with a config.json file in it.

python model.py $expNo$ 
(Example: python model.py 6)

Step 3: Run Inference

python model.py $expNo$ $sent.txt$ 
(Example: python model.py 6 sents.txt)

Step 4: Evaluate ROUGE scores

To evaluate for rouge, we use files2rouge, which itself uses pythonrouge.

Installation instructions:

pip install git+https://github.com/tagucci/pythonrouge.git
git clone https://github.com/pltrdy/files2rouge.git
cd files2rouge
python setup_rouge.py
python setup.py install

To run evaluation, simply run:

files2rouge summaries.txt references.txt

Cite

@inproceedings{malireddy2020scar,
  title={SCAR: Sentence Compression using Autoencoders for Reconstruction},
  author={Malireddy, Chanakya and Maniar, Tirth and Shrivastava, Manish},
  booktitle={Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics: Student Research Workshop},
  pages={88--94},
  year={2020}
}