This repository contains the code for our ACL 2020 paper:

Composing Elementary Discourse Units in Abstractive Summarization.

If you use this code, please cite our paper:

@inproceedings{li-etal-2020-composing,
    title = "Composing Elementary Discourse Units in Abstractive Summarization",
    author = "Li, Zhenwen  and
      Wu, Wenhao  and
      Li, Sujian",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020"
}

This project is based on Chen's work.

• Python 3 (tested on python 3.6)
• PyTorch 1.10.0
  • with GPU and CUDA enabled installation (though the code is runnable on CPU, it would be way too slow)
• Tensorflow 1.5.0
  • This is used for EDU segmentation.
• gensim
• cytoolz
• spacy
• allennlp
• tensorboardX
• pyrouge (for evaluation)

We can not provide the segmented data due to the copyright. If you want to train or evaluate our model on CNN/DailyMail dataset, you need to do the next several steps to get the dataset.

• Please follow the instructions here for downloading and preprocessing the CNN/DailyMail dataset.
• Segment the articles into EDUs by:

cd segment/src && python3 run.py --input_dir=[path/to/cnn&dailymail_finished_dir]

We use the code of Toward Fast and Accurate Neural Discourse Segmentation and modify to meet the structure of CNN/DailyMail dataset. Note this step may cost several hours to several days because we need to segment nearly 300,000 of articles.

• After segmenting the articles, there is a folder named segmented that contains three folders: train, val and test. Set the enviroment variable

export DATA=[path/to/segmented_dir]

• Get the labels of data by

python3 make_extraction_labels.py

Setup the official ROUGE packages at here. Please specify the path to your ROUGE package by setting the environment variable

export ROUGE=[path/to/rouge/directory]

After preparing the segmented and labeled data, yuo can train an EDUSum model yourself:

• pretrain the word2vec word embedding:

python train_word2vec.py --path=[path/to/word2vec]

• pretrain the EDU fusion module and EDU Selection module seperately:

python3 train_abstractor.py --path=[path/to/abstractor/model] --w2v=[path/to/word2vec/word2vec.128d.226k.bin]
python3 train_extractor_ml.py --path=[path/to/extractor/model] --w2v=[path/to/word2vec/word2vec.128d.226k.bin]

• Train the whole model end-to-end by RL:

python3 train_full_rl.py --path=[path/to/save/model] --abs_dir=[path/to/EDU_fusion_module] --ext_dir=[path/to/EDU_Selection_module]

To evaluate the trained model and generate summaries, run:

python3 decode_full_model.py --path=[path/to/save/decoded/files] --model_dir=[path/to/pretrained] --beam=[beam_size] [--test/--val]

To compute ROUGE score, firstly run:

python3 make_eval_references.py

to generate the reference files and then run

python3 eval_full_model.py --[rouge/meteor] --decode_dir=[path/to/save/decoded/files]