Dataset Card for "shaj"

Table of Contents

• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information
  • Contributions

Dataset Description

• Homepage:
• Repository: https://figshare.com/articles/dataset/SHAJ_Albanian_hate_speech_abusive_language/19333298/1
• Paper: https://arxiv.org/abs/2107.13592
• Point of Contact: Leon Derczynski
• Size of downloaded dataset files: 769.21 KiB
• Size of the generated dataset: 1.06 MiB
• Total amount of disk used: 1.85 MiB

Dataset Summary

This is an abusive/offensive language detection dataset for Albanian. The data is formatted following the OffensEval convention, with three tasks:

• Subtask A: Offensive (OFF) or not (NOT)
• Subtask B: Untargeted (UNT) or targeted insult (TIN)
• Subtask C: Type of target: individual (IND), group (GRP), or other (OTH)

Notes on the above:

• The subtask A field should always be filled.
• The subtask B field should only be filled if there's "offensive" (OFF) in A.
• The subtask C field should only be filled if there's "targeted" (TIN) in B.

The dataset name is a backronym, also standing for "Spoken Hate in the Albanian Jargon"

See the paper https://arxiv.org/abs/2107.13592 for full details.

Supported Tasks and Leaderboards

• Task A leaderboard at paperswithcode.com/sota/hate-speech-detection-on-shaj

Languages

Albanian (bcp47:sq-AL)

Dataset Structure

Data Instances

shaj

• Size of downloaded dataset files: 769.21 KiB
• Size of the generated dataset: 1.06 MiB
• Total amount of disk used: 1.85 MiB

An example of 'train' looks as follows.

{
  'id': '0', 
  'text': 'PLACEHOLDER TEXT', 
  'subtask_a': 1, 
  'subtask_b': 0, 
  'subtask_c': 0
}

Data Fields

• id: a string feature.
• text: a string.
• subtask_a: whether or not the instance is offensive; 0: OFF, 1: NOT
• subtask_b: whether an offensive instance is a targeted insult; 0: TIN, 1: UNT, 2: not applicable
• subtask_c: what a targeted insult is aimed at; 0: IND, 1: GRP, 2: OTH, 3: not applicable

Data Splits

Dataset Creation

Curation Rationale

Collecting data for enabling offensive speech detection in Albanian

Source Data

Initial Data Collection and Normalization

The text is scraped from comments on popular Albanian YouTube and Instagram accounts. An extended discussion is given in the paper in section 3.2.

Who are the source language producers?

People who comment on a selection of high-activity Albanian instagram and youtube profiles.

Annotations

Annotation process

The annotation scheme was taken from OffensEval 2019 and applied by two native speaker authors of the paper as well as their friends and family.

Who are the annotators?

Albanian native speakers, male and female, aged 20-60.

Personal and Sensitive Information

The data was public at the time of collection. No PII removal has been performed.

Considerations for Using the Data

Social Impact of Dataset

The data definitely contains abusive language.

Discussion of Biases

Other Known Limitations

Additional Information

Dataset Curators

The dataset is curated by the paper's authors.

Licensing Information

The authors distribute this data under Creative Commons attribution license, CC-BY 4.0.

Citation Information

@article{nurce2021detecting,
  title={Detecting Abusive Albanian},
  author={Nurce, Erida and Keci, Jorgel and Derczynski, Leon},
  journal={arXiv preprint arXiv:2107.13592},
  year={2021}
}

Contributions

Author-added dataset @leondz