Dataset Card for OffendES

Dataset Summary

Focusing on young influencers from the well-known social platforms of Twitter, Instagram, and YouTube, we have collected a corpus composed of Spanish comments manually labeled on offensive pre-defined categories. From the total corpus, we selected 30,416 posts to be publicly published, they correspond to the ones used in the MeOffendES competition at IberLEF 2021. The posts are labeled with the following categories:

• Offensive, the target is a person (OFP). Offensive text targeting a specific individual.
• Offensive, the target is a group of people or collective (OFG). Offensive text targeting a group of people belonging to the same ethnic group, gender or sexual orientation, political ideology, religious belief, or other common characteristics.
• Non-offensive, but with expletive language (NOE). A text that contains rude words, blasphemes, or swearwords but without the aim of offending, and usually with a positive connotation.
• Non-offensive (NO). Text that is neither offensive nor contains expletive language

Supported Tasks and Leaderboards

This dataset is intended for multi-class offensive classification and binary offensive classification. Competition MeOffendES task on offensive detection for Spanish at IberLEF 2021

Languages

• Spanish

Dataset Structure

Data Instances

For each instance, there is a string for the id of the tweet, a string for the emotion class, a string for the offensive class, and a string for the event. See the to explore more examples.

{'comment_id': '8003',
 'influencer': 'dalas',
 'comment': 'Estupido aburrido',
 'label': 'NO',
 'influencer_gender': 'man',
 'media': youtube
 }

Data Fields

• comment_id: a string to identify the comment
• influencer: a string containing the influencer associated with the comment
• comment: a string containing the text of the comment
• label: a string containing the offensive gold label
• influencer_gender: a string containing the genre of the influencer
• media: a string containing the social media platform where the comment has been retrieved

Data Splits

The OffendES dataset contains 3 splits: train, validation, and test. Below are the statistics for each class.

OffendES	Number of Instances in Split per class
Class	Train	Validation	Test
NO	13,212	64	9,651
NOE	1,235	22	2,340
OFP	2,051	10	1,404
OFG	212	4	211
Total	16,710	100	13,606

Dataset Creation

Source Data

Twitter, Youtube, Instagram

Who are the annotators?

Amazon Mechanical Turkers

Additional Information

Licensing Information

The OffendES dataset is released under the Apache-2.0 License.

Citation Information

@inproceedings{plaza-del-arco-etal-2021-offendes,
    title = "{O}ffend{ES}: A New Corpus in {S}panish for Offensive Language Research",
    author = "{Plaza-del-Arco}, Flor Miriam  and Montejo-R{\'a}ez, Arturo and Ure{\~n}a-L{\'o}pez, L. Alfonso  and Mart{\'\i}n-Valdivia, Mar{\'\i}a-Teresa",
    booktitle = "Proceedings of the 12th Language Resources and Evaluation Conference",
    month = sep,
    year = "2021",
    address = "Held Online",
    url = "https://aclanthology.org/2021.ranlp-1.123.pdf",
    language = "English",
    pages = "1096--1108"
}

@article{meoffendes2021,
  title="{{Overview of MeOffendEs at IberLEF 2021: Offensive Language Detection in Spanish Variants}}",
  author="{Flor Miriam Plaza-del-Arco and Casavantes, Marco and Jair Escalante, Hugo and Martín-Valdivia, M. Teresa and Montejo-Ráez, Arturo and {Montes-y-Gómez}, Manuel and Jarquín-Vásquez, Horacio and Villaseñor-Pineda, Luis}",
  journal="Procesamiento del Lenguaje Natural",
  url = "https://bit.ly/3QpRDfy",
  volume="67",
  pages="183--194",
  year="2021"
}