README.md

metadata

task_categories:
  - text-classification
language:
  - en
size_categories:
  - 1K<n<10K
extra_gated_prompt: >-
  This dataset includes sensitive and explicit material.  You may encounter 
  descriptions that are sexually explicit, graphic, or otherwise unsuitable for
  all audiences. Viewer discretion is strongly advised. Please proceed only if
  you are comfortable and consent to viewing this type of content. You agree to
  not use the dataset to conduct experiments that cause harm to human subjects.
extra_gated_fields:
  Company: text
  Country: country
  I want to use this dataset for:
    type: select
    options:
      - Research
      - Education
      - label: Other
        value: other
  I agree to use this dataset for non-commercial use ONLY: checkbox
license: cc-by-2.0

X-Sensitive is a multi-label dataset designed to identify sensitive language in social media. It consists of 7 labels and includes a total of 8,000 posts extracted from X. Each post is assigned one or more of the following labels based on its content: Drugs, Sex, Conflictual, Spam, Profanity, and Self-harm. More details in the reference paper.

The goal of X-Sensitive is to serve as a valuable resource for developing online moderation tools. The following models have been trained on X-Sensitive with this aim:

• twitter-roberta-large-sensitive-multilabel
• twitter-roberta-base-sensitive-multilabel

We also provide binary versions of the models, where each post is classified as either sensitive or not-sensitive:

• twitter-roberta-large-sensitive-binary
• twitter-roberta-base-sensitive-binary

Dataset Structure

Data Splits

Name	#Entries
train	5,000
test	2,000
validation	1,000

Data Instances

An example of train looks as follows.

{'#labels': 1,
 'conflictual': 0,
 'conflictual_highlight': array([], dtype=object),
 'drugs': 0,
 'drugs_highlight': array([], dtype=object),
 'keyword': 'fuckin',
 'labels': array(['profanity'], dtype=object),
 'profanity': 1,
 'profanity_highlight': array([array(['fucking'], dtype=object), array(['fucking'], dtype=object),
       array(['fucking'], dtype=object)], dtype=object),
 'selfharm': 0,
 'selfharm_highlight': array([], dtype=object),
 'sex': 0,
 'sex_highlight': array([], dtype=object),
 'spam': 0,
 'spam_highlight': array([], dtype=object),
 'text': 'i think the idea of aliens is so fucking cool'}

Labels

Label Number	Label Name	Description
0	conflictual	Conflictual language. An attack based on protected (race, color, caster, gender, etc) or other categories.
1	profanity	Language containing slurs and profanity even if they are not directed towards a specific entity.
2	sex	Sexually Explicit Content. Pornographic or other types of sexual content.
4	selfharm	Self-harm. Posts depicting, promoting, or glorifying violence or harm against oneself such as eating disorders or suicide.
5	spam	Irrelevant content that is unsolicited.

Citation Information

@article{antypas2024sensitive,
  title={Sensitive Content Classification in Social Media: A Holistic Resource and Evaluation},
  author={Antypas, Dimosthenis and Sen, Indira and Perez-Almendros, Carla and Camacho-Collados, Jose and Barbieri, Francesco},
  journal={arXiv preprint arXiv:2411.19832},
  year={2024}
}