Skip to content
Filter script for JNC/JAMUL
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.


Type Name Latest commit message Commit time
Failed to load latest commit information.


This is a repository for headline Generation of multi-lengths Outpus in Japanese (Gingo). We conviniently call this task Gingo. In this repository, we provide preprocess scripts of JAMUL/JNC corpus and evaluation scripts for Japanese summarization in ROUGE metric.


JApanese MUlti-Length Headline Corpus (JAMUL) is a corpus containing 1,524 news articles and their length-sensitive headlines in 10, 13, and 26 characters long for digital media and their length-insensitive headlines for paper. The articles and headlines were all published between September 2017 and March 2018.

JAMUL.csv is arranged in order of per line as below.

  • article
  • paper headline
  • 26 characters headline
  • 13 characters headline
  • 10 characters headline

You can get JAMUL.csv by sending e-mail to media-lab-rndrpr[atmark] (please replace [atmark] to @).


  • Python 3.4+
  • mecab-python3
  • pythonrouge

We have a requirements.txt file for installing them:

pip -r requirments.txt

JAMUL preprocess script

  • filter script to create the same test set in our paper.
python --input_path ./JAMUL.csv --output_path ./testset.csv

JNC Corpus

Japanese News Corpus (JNC) is a collection of 1,829,231 pairs of the three lead sentences of articles and their print headlines published from 2007 to 2016. We use this dataset to train our seq2seq model. You can can get JNC corpus for a fee (more details ).

JNC preprocess script

  • filter script to create the same training data in our paper.
python --input_path ./JNC-corpus.json --output_path ./output_dir

ROUGE evaluation

  • ROUGE evaluation script

We used the following options for in case of specifying 26 characters long.

python --reference /path/to/test_headlines_of_JAMUL --predict /path/to/generated_headlines --trim 26
You can’t perform that action at this time.