metadata

task_categories:
  - text-classification
  - text-generation
language:
  - en
tags:
  - not-for-all-audiences
  - legal
pretty_name: dynamically generated hate speech dataset

Dataset Card for dynamically generated hate speech dataset

Dataset Description

Homepage: GitHub
Point of Contact: [email protected]

Dataset Summary

This is a copy of the Dynamically-Generated-Hate-Speech-Dataset, presented in this paper by

Bertie Vidgen, Tristan Thrush, Zeerak Waseem and Douwe Kiela

Original README from GitHub

Dynamically-Generated-Hate-Speech-Dataset

ReadMe for v0.2 of the Dynamically Generated Hate Speech Dataset from Vidgen et al. (2021). If you use the dataset, please cite our paper in the Proceedings of ACL 2021, and available on Arxiv. Contact Dr. Bertie Vidgen if you have feedback or queries: [email protected].

The full author list is: Bertie Vidgen (The Alan Turing Institute), Tristan Thrush (Facebook AI Research), Zeerak Waseem (University of Sheffield) and Douwe Kiela (Facebook AI Research). This paper is an output of the Dynabench project: https://dynabench.org/tasks/5#overall

Dataset descriptions

v0.2.2.csv is the full dataset used in our ACL paper.

v0.2.3.csv removes duplicate entries, all of which occurred in round 1. Duplicates come from two sources: (1) annotators entering the same content multiple times and (2) different annotators entering the same content. The duplicates are interesting for understanding the annotation process, and the challenges of dynamically generating datasets. However, they are likely to be less useful for training classifiers and so are removed in v0.2.3. We did not lower case the text before removing duplicates as capitalisations contain potentially useful signals.

Overview

The Dynamically Generated Hate Speech Dataset is provided in one table.

'acl.id' is the unique ID of the entry.

'Text' is the content which has been entered. All content is synthetic.

'Label' is a binary variable, indicating whether or not the content has been identified as hateful. It takes two values: hate, nothate.

'Type' is a categorical variable, providing a secondary label for hateful content. For hate it can take five values: Animosity, Derogation, Dehumanization, Threatening and Support for Hateful Entities. Please see the paper for more detail. For nothate the 'type' is 'none'. In round 1 the 'type' was not given and is marked as 'notgiven'.

'Target' is a categorical variable, providing the group that is attacked by the hate. It can include intersectional characteristics and multiple groups can be identified. For nothate the type is 'none'. Note that in round 1 the 'target' was not given and is marked as 'notgiven'.

'Level' reports whether the entry is original content or a perturbation.

'Round' is a categorical variable. It gives the round of data entry (1, 2, 3 or 4) with a letter for whether the entry is original content ('a') or a perturbation ('b'). Perturbations were not made for round 1.

'Round.base' is a categorical variable. It gives the round of data entry, indicated with just a number (1, 2, 3 or 4).

'Split' is a categorical variable. it gives the data split that the entry has been assigned to. This can take the values 'train', 'dev' and 'test'. The choice of splits is explained in the paper.

'Annotator' is a categorical variable. It gives the annotator who entered the content. Annotator IDs are random alphanumeric strings. There are 20 annotators in the dataset.

'acl.id.matched' is the ID of the matched entry, connecting the original (given in 'acl.id') and the perturbed version.

For identities (recorded under 'Target') we use shorthand labels to constructed the dataset, which can be converted (and grouped) as follows:

  none -> for non hateful entries 
  NoTargetRecorded -> for hateful entries with no target recorded
  
  mixed -> Mixed race background
  ethnic minority -> Ethnic Minorities
  indig -> Indigenous people
  indigwom -> Indigenous Women
  non-white -> Non-whites (attacked as 'non-whites', rather than specific non-white groups which are generally addressed separately)
  trav -> Travellers (including Roma, gypsies)

  bla -> Black people
  blawom -> Black women
  blaman -> Black men
  african -> African (all 'African' attacks will also be an attack against Black people)
  
  jew -> Jewish people
  mus -> Muslims
  muswom -> Muslim women

  wom -> Women	
  trans -> Trans people
  gendermin -> Gender minorities, 
  bis -> Bisexual
  gay -> Gay people (both men and women)
  gayman -> Gay men
  gaywom -> Lesbians	
  
  dis -> People with disabilities
  working -> Working class people
  old -> Elderly people

  asi -> Asians
  asiwom -> Asian women
  east -> East Asians
  south -> South Asians (e.g. Indians)
  chinese -> Chinese people
  pak -> Pakistanis
  arab -> Arabs, including people from the Middle East

  immig -> Immigrants
  asylum -> Asylum seekers
  ref -> Refguees
  for -> Foreigners
  
  eastern european -> Eastern Europeans
  russian -> Russian people
  pol -> Polish people
  hispanic -> Hispanic people, including latinx and Mexicans

  nazi -> Nazis ('Support' type of hate)
  hitler -> Hitler ('Support' type of hate)

Code

Code was implemented using hugging face transformers library.

Additional Information

Licensing Information

The original repository does not provide any license, but is free for use with proper citation of the original paper in the Proceedings of ACL 2021, available on Arxiv

Citation Information

cite as arXiv:2012.15761 or https://doi.org/10.48550/arXiv.2012.15761