The datasets contain article-headline pairs obtained from Japanese Wikinews. The articles and headlines are segmented to words using mecab-ipadic.

In this repository, there are following three version datasets according to the article length:

• full-articles: the dataset with articles more than 10 tokens, and headlines;
• long version: the dataset with articles extracted from the first five sentences or 256 tokens, and headlines.
• short version: the dataset with articles extracted from the three sentences or 128 tokens, and headlines.

Table1 Number of documents

Table2 N-gram overlaps in headline