Awesome Open Source
Awesome Open Source


This repository contains the implementation details and datasets used in On Generating Extended Summaries of Long Documents paper at the AAAI-21 Workshop on Scientific Document Understanding (SDU 2021).

Conda environment: preliminary setup

To install the required packages, please run conda yml file that you find in the root directory using the following command:

conda env create -f environment.yml

How to run...

IMPORTANT: The following commands should be run under src/ directory.


To start with, you first need to download the datasets that are intended to work with the code base. You can download them from following links:

Dataset Download Link
arXiv-Long Download
PubMed-Long Download

After downloading the dataset, you will need to uncompress it using the following command:

tar -xvf pubmedL.tar.gz 

This will uncompress the pubmedL tar file into the current directory. The directory will include the single json files of different sets including training, validation, and test.

FORMAT Each paper file is structured within a a json object with the following keys:

  • "id" (String): the paper ID
  • "abstract" (String): the abstract text of the paper. This field is different from "gold" field for the datasets that have different ground-truth than the abstract.
  • "gold" (List <List<>>): the ground-truth summary of the paper, where the inner list is the tokens associated with each gold summary sentence.
  • "sentences" (List <List<>>): the source sentences of the full-text. The inner list contains 5 indices, each of which represents different fields of the source sentence:
    • Index [0]: tokens of the sentences (i.e., list of tokens).
    • Index [1]: textual representation of the section that the sentence belongs to.
    • Index [2]: Rouge-L score of the sentence with the gold summary.
    • Index [3]: textual representation of the sentences.
    • Index [4]: oracle label associated with the sentence (0, or 1).
    • Index [5]: the section id assigned by sequential sentence classification package. For more information, please refer to this repository

Preparing Data

Simply run the bash script with providing the dataset directory. This script will use two functions to first create aggregated json files, and then preparing them for pretrained language models' usage.

Please note that if you want to use your custom dataset and create torch files, you will need to frame the format of your dataset to the given format in the Dataset section.


The full training scripts are inside bash file. To run it on your machine, you will need to change the directories to fit in your needs:



# Specifiying GPUs either single GPU, or multi-GPU

# You don't need to modify these below 
LOG_DIR=../logs/$(echo $MODEL_PATH | cut -d \/ -f 6).log
mkdir -p ../results/$(echo $MODEL_PATH | cut -d \/ -f 6)
RESULT_PATH_TEST=../results/$(echo $MODEL_PATH | cut -d \/ -f 6)/




The inference scripts are inside bash file. To run it on your machine, you will need to modify the file directories:

# path to the data directory

# path to the trained model directory

# path to the best trained model (or the checkpoint that you want to run inference on)

# GPU machines, either multi or single GPU




If you plan to use this work, please cite the following papers:

  title={On Generating Extended Summaries of Long Documents},
  author={Sajad Sotudeh and Arman Cohan and Nazli Goharian},
  booktitle={The AAAI-21 Workshop on Scientific Document Understanding (SDU 2021)},
  title={GUIR @ LongSumm 2020: Learning to Generate Long Summaries from Scientific Documents},
  author={Sajad Sotudeh and Arman Cohan and Nazli Goharian},
  booktitle={First Workshop on Scholarly Document Processing (SDP 2020)},

Get A Weekly Email With Trending Projects For These Topics
No Spam. Unsubscribe easily at any time.
Python (1,124,836
Dataset (3,540
Text Summarization (217
Related Projects