readerbench/news-ro-offense

a novel Romanian language dataset for offensive message detection with manually annotated comment from a local Romanian news website (stiri de cluj) into five classes

News-RO-Offense

a novel Romanian language dataset for offensive message detection with manually annotated comment from a local Romanian news website (stiri de cluj) into five classes:

• non-offensive
• targeted insults
• racist
• homophobic
• sexist

Resulting in 4052 annotated messages

Corpus statistics

Class	Label	# examples	Percentage
Non-offensive	0	2682	66.19%
Targeted insult	1	777	19.18%
Racist	2	252	6.22%
Homophobic	3	186	4.59%
Sexist	4	155	3.82%
TOTAL		4052

Data structure

comments.csv

Csv containing all annotated comments (Fields comment_text and LABEL)

additionally, all comments have the article_id for the originating article and, if they were a reply to another comment, the reply-to comment_id

articles.csv

contains all 160 articles that were scraped for comments. The available fields are - Publishing Date, Article Title, Article Tags list and Article content

Paper citation

@inproceedings{cojocaru2022news-ro-offense,
  title={News-RO-Offense - A Romanian Offensive Language Dataset and Baseline Models Centered on News Article Comments},
  author={Cojocaru, Andreea and Paraschiv, Andrei and Dascalu, Mihai},
  booktitle={Proceedings of the international conference on human-computer interaction-RoCHI 2022},
  pages={--},
  year={2022}
}