This repository contains data and code for the NAACL 2018 paper "A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents". Please note that the code is not actively maintained.
Two datasets of long and structured documents (scientific papers) are provided. The datasets are obtained from ArXiv and PubMed OpenAccess repositories.
ArXiv dataset: Download
PubMed dataset: Download
The datasets are rather large. You need about 5G disk space to download and about 15G additional space when extracting the files. Each tar file consists of 4 files. train.txt, val.txt, test.txt respectively correspond to the training, validation, and test sets. The vocab file is a plaintext file for the vocabulary.
The files are in jsonlines format where each line is a json object corresponding to one scientific paper from ArXiv or PubMed. The abstract, sections and body are all sentence tokenized. The json objects are in the following format:
{
'article_id': str, {
'abstract_text': List[str], 'article_id': str,
'article_text': List[str], 'methodology': List[str],
'section_names': List[str], 'method_summary': List[str]
'sections': List[List[str]] }
}
| Dataset | # docs | avg. doc. length (words) | avg. summary length (words) |
|---|---|---|---|
| CNN | 92K | 656 | 43 |
| Daily Mail | 219K | 693 | 52 |
| NY Times | 655K | 530 | 38 |
| Gigaword | 4M | 31 | 8 |
| Pubmed | 133K | 3016 | 203 |
| arXiv | 215K | 4938 | 220 |
| arXiv-Methodology (this work) | 1006 | 1358 | 75 |
| RG-1 | RG-2 | RG-3 | RG-L |
|---|---|---|---|
| 24.54 | 3.08 | 0.37 | 19.93 |
we here calibrate gaussians in isotropic neighborhood and gaussians , in order to their variability in aluminum in their energies . the dominant model is represented , and the latter , together with the cme : the multi - cme which
in this paper , we present a method to quantify the hamming - central central narrowing ( tf ) scenario of the half - processing chain , which nearest - contacted of m31 the m31 and other produced from the strength distance . in the model we model the numbers uncertainties using the best - field model and the decay of the respective satellite . furthermore , in which the significance of m31 s , the strong dirac cone exists from the coordinate distance and the actual earth torino into account captured with the mimes .
we present the theoretical space for the network rate of the network of cortical trains . we have calculated a two - range of the network of network in a network of inhibition - spike potential , and the sub - ray emitting poisson trains .
The dataset is also available on Tensorflow Datasets which makes it easy to use within Tensorflow or colab.
The code is based on the pointer-generator network code by See et al. (2017). Refer to their repo for documentation about the structure of the code.
You will need python 3.6 and Tensorflow 1.5 to run the code. The code might run with later versions of Tensorflow but it is not tested. Checkout other dependencies in requirements.txt file. A small sample of the dataset is already provided in this repo. To run the code with the sample data unzip the files in the data directory and simply execute the run script: ./run.sh. To train the model with the entire dataset, first convert the jsonlines files to binary using the the following script: scripts/json_to_bin.py and modify the corresponding training data path in the run.sh script.
If you ended up finding this paper or repo useful please cite:
"A Discourse-Aware Attention Model for Abstractive Summarization of Long Documents"
Arman Cohan, Franck Dernoncourt, Doo Soon Kim, Trung Bui, Seokhwan Kim, Walter Chang, and Nazli Goharian
NAACL-HLT 2018
Another relevant reference is Pointer-Generator network by See et al. (2017):
"Get to the point: Summarization with pointer-generator networks."
Abigail See, Peter J. Liu, and Christopher D. Manning.
ACL (2017).